Robot control method, device and robot

Abstract

The present disclosure relates to the field of artificial intelligence and the field of robots, and provides a control method and device of a robot and the robot, the method comprising: obtaining a plurality of research paradigms according to a research task for a robot algorithm; iteratively performing research based on the plurality of research paradigms to obtain a target algorithm corresponding to the research task; and controlling the robot to perform a task based on the target algorithm, wherein in the process of iteratively performing research, when research is performed based on at least one research paradigm in the plurality of research paradigms, research results obtained based on other research paradigms are reused.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and robotics, and specifically to a robot control method, device, and robot. Background Technology

[0002] With the development of artificial intelligence and robotics technologies, various research paradigms have been proposed for robot algorithms. These paradigms enable the construction, training, and optimization of robot algorithms, and the resulting robot algorithms can be deployed on real robots to guide them in performing tasks.

[0003] In some cases, Large Language Models (LLMs) can be used to study robot algorithms based on these research paradigms, resulting in the final algorithm used to control the robot. However, due to the differences in research focus and achieved effects among various research paradigms, robot algorithms based on a single paradigm may exhibit uneven performance when deployed on robots, making it difficult to comprehensively improve the robot's task execution quality and capabilities. Summary of the Invention

[0004] This disclosure provides a robot control method, apparatus, and robot to at least solve the aforementioned problems in the related art.

[0005] According to a first aspect of this disclosure, a robot control method is provided, the control method comprising: acquiring multiple research paradigms based on a research task for a robot algorithm, wherein each research paradigm refers to a methodological system for researching robot algorithms; iteratively performing research based on the multiple research paradigms to obtain a target algorithm corresponding to the research task; and controlling the robot to perform the task based on the target algorithm, wherein, during the iterative execution of the research, when performing research based on at least one of the multiple research paradigms, research results obtained based on other research paradigms are reused.

[0006] Optionally, in each iteration, the research is performed as follows: based on the research paradigm of this iteration and the accumulated research result set, the research results of this iteration are obtained, wherein in the first iteration, the accumulated research result set is a preset set; and / or, in response to the current research progress meeting the paradigm switching condition, the research results of this iteration are added to the accumulated research result set, another research paradigm among the multiple research paradigms is used as the research paradigm for the next iteration, and the next iteration is executed; and / or, in response to the current research progress not meeting the paradigm switching condition and not meeting the real machine migration condition, the research results of this iteration are added to the accumulated research result set, the research paradigm of this iteration is used as the research paradigm for the next iteration, and the next iteration is executed; in response to the current research progress not meeting the paradigm switching condition but meeting the real machine migration condition, the target algorithm is obtained based on the research results of this iteration.

[0007] Optionally, the paradigm shift condition includes at least one of the following: identifying a dominant paradigm among the plurality of research paradigms that is suitable for the current research progress; or the budget allocated to the research paradigm for this iteration has been exhausted.

[0008] Optionally, obtaining the research results of this iteration based on the research paradigm of this iteration and the accumulated set of research results includes: using a large language model, generating an algorithm experiment package based on the research paradigm of this iteration and the accumulated set of research results, wherein the algorithm experiment package includes data for performing robot algorithm experiments; and performing robot algorithm experiments based on the algorithm experiment package to obtain the research results of this iteration.

[0009] Optionally, the step of executing robot algorithm experiments based on the algorithm experiment package to obtain the research results of this iteration includes: performing security verification on the generated algorithm experiment package; intercepting the algorithm experiment package in response to the algorithm experiment package failing the security verification; and / or, in response to the algorithm experiment package passing the security verification, executing robot algorithm experiments in a runtime container based on the algorithm experiment package to obtain the research results of this iteration, wherein the algorithm running in the runtime container is isolated from the outside of the runtime container.

[0010] Optionally, the step of using another research paradigm among the plurality of research paradigms as the research paradigm for the next iteration and executing the next iteration includes: The strategy object of the current research paradigm is replaced by the strategy object of the other research paradigm, while keeping the storage handle of the accumulated research results set unchanged, so that the other research paradigm executes the next iteration based on the accumulated research results set before the switch.

[0011] Optionally, controlling the robot to perform a task based on the target algorithm includes: generating control instructions for controlling the robot to perform a task based on the target algorithm; triggering an emergency stop of the robot in response to the control instructions exceeding a preset limit range; and controlling the robot to perform a task based on the control instructions in response to the control instructions not exceeding the limit range, wherein the limit range is used to restrict instructions in the control instructions related to the physical parameters of the robot.

[0012] Optionally, controlling the robot to perform a task based on the target algorithm includes: performing at least one algorithm optimization for the target algorithm to obtain an optimized target algorithm; controlling the robot to perform a task based on the optimized target algorithm, wherein each algorithm optimization includes: controlling the robot to perform a task based on the algorithm to be optimized at each speed level according to a preset number of speed levels, from low to high, to obtain real machine data, wherein in the first optimization, the algorithm to be optimized is the target algorithm; optimizing the target algorithm based on the real machine data to obtain the optimized algorithm for this time; in response to the optimized algorithm satisfying a preset output condition, using the optimized algorithm for this time as the optimized target algorithm; in response to the optimized algorithm not satisfying the preset output condition, using the optimized algorithm for this time as the algorithm to be optimized for the next optimization and performing the next optimization.

[0013] Optionally, the step of controlling the robot to perform tasks based on an algorithm to be optimized at each preset speed level, from low to high, to obtain real machine data, includes: adjusting the algorithm to be optimized in response to the robot failing to perform a task at any speed level, and controlling the robot to perform the failed task based on the adjusted algorithm until all tasks at the current speed level are successfully performed, and then switching to the next speed level; and using the data generated during the process of performing tasks at all speed levels as the real machine data in response to the robot successfully performing a task at the highest speed level.

[0014] According to a second aspect of this disclosure, a robot control device is provided, the control device comprising: an acquisition unit configured to acquire a plurality of research paradigms based on a research task for a robot algorithm, wherein each research paradigm refers to a methodological system for researching robot algorithms; a research unit configured to iteratively perform research based on the plurality of research paradigms to obtain a target algorithm corresponding to the research task; and a control unit configured to control the robot to perform the task based on the target algorithm, wherein, during the iterative execution of the research, when performing research based on at least one of the plurality of research paradigms, research results obtained based on other research paradigms are reused.

[0015] According to a third aspect of this disclosure, an electronic device is provided, the electronic device comprising: a processor; and a memory for storing processor-executable instructions, wherein the processor-executable instructions, when executed by the processor, cause the processor to perform a robot control method according to this disclosure.

[0016] According to a fourth aspect of this disclosure, a robot is provided that includes electronic equipment according to this disclosure, or the robot is communicatively connected to electronic equipment according to this disclosure.

[0017] According to a fifth aspect of this disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform a robot control method according to this disclosure.

[0018] According to a sixth aspect of this disclosure, a computer program product is provided, including computer-executable instructions that, when executed by at least one processor, implement the robot control method according to this disclosure.

[0019] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects: By adopting the robot control scheme according to this disclosure, research can be carried out iteratively based on multiple research paradigms for the same research task, and at least one research paradigm can reuse the research results of other research paradigms, eliminating information barriers between paradigms and realizing cross-paradigm result sharing. This allows the final robot algorithm to integrate the advantages of each paradigm. Controlling the robot to perform tasks based on such an algorithm can improve the robot's task execution quality and capabilities.

[0020] It will be understood that the above general description and the following detailed description are merely examples and do not limit this disclosure. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0022] Figure 1 This is a schematic flowchart illustrating a robot control method according to an exemplary embodiment of the present disclosure.

[0023] Figure 2 This is a schematic flowchart illustrating an iterative execution study according to an exemplary embodiment of the present disclosure.

[0024] Figure 3This is a schematic flowchart illustrating an orchestration multi-paradigm unification study according to exemplary embodiments of the present disclosure.

[0025] Figure 4 This is a schematic flowchart illustrating a five-stage bridging process during algorithm migration according to an exemplary embodiment of the present disclosure.

[0026] Figure 5 This is a schematic diagram illustrating an example architecture of two-layer asymmetric security control according to exemplary embodiments of the present disclosure.

[0027] Figure 6 This is a schematic flowchart illustrating an application example of a robot control method according to an exemplary embodiment of the present disclosure.

[0028] Figure 7 This is a schematic diagram illustrating an example architecture for implementing a control method for a robot according to an exemplary embodiment of the present disclosure.

[0029] Figure 8 This is a schematic block diagram of a robot control device according to exemplary embodiments of the present disclosure.

[0030] Figure 9 This is a block diagram of an electronic device according to exemplary embodiments of the present disclosure. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0032] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0033] It should be noted that the phrase "at least one of several items" in this disclosure refers to three parallel cases: "any one of the several items", "a combination of any number of the several items", and "all of the several items". For example, "including at least one of A and B" includes the following three parallel cases: (1) including A; (2) including B; (3) including A and B. Another example is "performing at least one of step one and step two", which means the following three parallel cases: (1) performing step one; (2) performing step two; (3) performing both step one and step two.

[0034] It should also be noted that the data collected in the embodiments of this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0035] To better understand, the following is a definition of the terms used in this article: Research Orchestrator: The intelligent core that unifies the scheduling of multiple paradigms and four roles; Research paradigm: A methodological framework used to study robot algorithms, such as evolutionary loops, best-priority tree search, and minimal fast loops; Shared Experiment History Database (Experiment DB): A persistent database of experimental nodes shared across paradigms; Cognition Store: A cross-paradigm shared domain knowledge base; Multi-stage Sim2Real bridge: Five-stage pipeline migration; Manual approval gate: A mandatory approval process at the architecture layer that cannot be bypassed before accessing the actual device; Physical hard-limit layer: A security layer that forcibly truncates control commands before they are issued; Domain randomization (ADR): Randomization of simulation parameters; Large Language Model / Visual Language Model (LLM / VLM): A general model for hypothesis / code / analysis / visual analysis.

[0036] As mentioned earlier, in robot control, robot algorithms based on a single paradigm exhibit uneven task performance after being deployed to robots, making it difficult to comprehensively improve the robot's task execution quality and capabilities.

[0037] Specifically, in some schemes, a "hypothesis-experiment-analysis-conclusion" process can be driven by best-priority tree search, in which a visual language model can be introduced to examine the charts. However, in such a process, only the tree search research paradigm is supported.

[0038] Some proposals integrate evolutionary computation, LLM reasoning, and knowledge retrieval, employing a research framework that combines dual-memory systems, MAP-Elites sampling of UCB1 and Island models, and a three-role division of labor. However, such proposals only support one research paradigm: evolutionary cycles.

[0039] In some schemes, algorithm research can be carried out with a minimalist loop of "single file editing, fixed time budget, save and rollback" as the core. However, it only supports this minimalist loop research paradigm.

[0040] In some approaches, LLM can be used to perform evolutionary optimization on the robot's reinforcement learning reward code and automatically evaluate it in simulation. Alternatively, LLM can be used to generate reward functions and automatically randomized domain parameters based on task and safety instructions to complete the migration from simulation to real machine. However, LLM can only perform evolutionary optimization and simulation based on a fixed research paradigm.

[0041] The aforementioned existing technologies provide readily available and mature methods (UCB1, MAP-Elites, best-priority tree search, ADR, reward code evolution, container sandbox, etc.).

[0042] In the above-mentioned schemes, the research on robot algorithms is carried out under a single research paradigm, and the infrastructure is not shared. Different paradigms have their own applicable scenarios (for example, tree search is suitable for deep ablation, evolution is suitable for architecture / policy search, and minimalism is suitable for rapid hyperparameter iteration, etc.). There is a lack of a scheme that can unify the orchestration of multiple paradigms. Furthermore, robot algorithms based on a single paradigm also show uneven performance of robot tasks after being deployed to robots, making it difficult to comprehensively improve the task execution quality and capabilities of robots.

[0043] Furthermore, in the process of migrating algorithms to real devices, only the migration accuracy or robustness is considered. The migration strategy is often "deploy immediately after training" or only constrained by the safety instructions in the prompt, lacking verification, security gating and feedback loop. As a result, the task execution on the real device after migration is at risk of reliability.

[0044] Furthermore, these solutions lack strict code control over the LLM-generated code that may be deployed to real devices, which poses risks to code execution and robot security after migration to real devices.

[0045] In view of the above problems, exemplary embodiments of the present disclosure provide a robot control method, robot control device, electronic device, robot, computer-readable storage medium, and computer program product that can solve or at least alleviate at least one of the above problems.

[0046] In a first aspect of the exemplary embodiments of this disclosure, a robot control method is provided, which can be applied to scenarios where the robot is controlled to perform tasks. The method can be executed on any electronic device, such as a tablet computer, laptop computer, digital assistant, wearable device, etc. However, this is only an example scenario. The method according to the exemplary embodiments of this disclosure can also be applied to other application scenarios. For example, a terminal can send a request to a server via a network, and the server can execute the robot control by performing the method according to the exemplary embodiments of this disclosure. Here, the server can be a standalone server, a server cluster, a cloud computing platform, or a virtualization center.

[0047] The robot control method according to the exemplary embodiments of this disclosure relates to the interdisciplinary field of artificial intelligence and robotics. It can be uniformly arranged across multiple research paradigms, safely completing research hypothesis generation, experiment execution, multi-dimensional evaluation, knowledge accumulation and iterative improvement in a closed loop composed of simulated and real physical robots. After simulation verification, the strategy is transferred to the physical robot for verification through multi-layer safety control. The robot can be controlled to perform tasks based on the final algorithm, thereby improving the robot's comprehensive performance and overall task execution effect.

[0048] The following will refer to Figure 1 The specific steps of a robot control method according to exemplary embodiments of the present disclosure are described. For example... Figure 1 As shown, the processing method may include the following steps: In step S110, multiple research paradigms can be obtained based on the research task for the robot algorithm.

[0049] Here, each research paradigm can refer to a methodological system used to study robotic algorithms, or in other words, a methodological system that defines how to conduct autonomous research on robotic algorithms (e.g., including but not limited to methods for generating research hypotheses, organizing experiments, and iterative optimization paths). For example, research paradigms may include evolutionary loops, best-priority tree search, and minimalist fast loops, with different paradigms differing in sampling strategies, expansion strategies, and / or iterative methods. Based on the selected research paradigm, robotic algorithms can be constructed, trained, and optimized, and the resulting robotic algorithms can be deployed to robots.

[0050] As an example, the aforementioned research paradigms can be selected and set according to actual needs. For example, they can include existing research paradigms (such as tree search, evolutionary cycle, minimal cycle, etc. mentioned above), or other arbitrary research paradigms. As an example, before executing this control method, candidate research paradigms for different research tasks can be prepared in advance. In this step S110, multiple corresponding research paradigms can be selected from the candidate research paradigms according to the current research task.

[0051] Here, the research task for the robot algorithm may include, but is not limited to, the construction, training, optimization, and updating of the algorithm. As an example, the research task may be input by a user, such as natural language text. For instance, the research task may include the training task of a robot algorithm with constraints on the number of model parameters or model performance. In the method of this disclosure, in response to receiving a research instruction containing the research task input by a user, research can be performed using a large model, such as a large language model and / or a visual language model, based on multiple research paradigms (e.g., generating hypotheses, writing experimental code, analyzing results and visually analyzing failure cases), to obtain a target robot algorithm that satisfies the constraints on the number of model parameters or model performance. Furthermore, the large language model and / or visual language model mentioned herein may employ general pre-trained large language models and visual language models; this disclosure does not impose any particular limitation on this.

[0052] Robotics algorithms can refer to algorithms that can be deployed on robots to guide them in performing tasks. Examples of robotic algorithms include, but are not limited to: perception algorithms for recognizing objects and / or locations around the robot; planning algorithms for planning task execution paths for the robot; and control algorithms for controlling the robot's joints and movements. For instance, robotic algorithms can include, but are not limited to: vision-language-action models, vision-language navigation models, manipulation strategies, motion control algorithms, and simulation-to-real-world transfer algorithms. Furthermore, robotic algorithms here can be artificial intelligence (AI) model algorithms or non-AI model algorithms, such as classical robotics algorithms.

[0053] Furthermore, in the embodiments disclosed herein, no particular limitations are placed on the specific type and application scenario of the robot. As an example, the robot may be a robot capable of deploying the aforementioned robot algorithms and performing tasks, such as including but not limited to: manipulators with fixed bases, such as industrial robotic arms and collaborative robotic arms; wheeled or tracked mobile robots; legged robots, such as quadruped robots and wheel-legged robots; humanoid robots with two arms and / or two legs; and composite robots composed of a mobile chassis and robotic arms.

[0054] As an example, the robot algorithm and the tasks performed by the robot can be determined according to the type of robot and the application scenario. For example, for an operational robot, the robot algorithm can be a vision-language-motion model and a grasping operation strategy, and the robot can perform grasping tasks; for a mobile robot, the robot algorithm can be a vision-language navigation model and a local path planning and motion control algorithm, and the robot can perform handling and delivery tasks; for a legged robot, the robot algorithm can be a gait generation and balance control motion control algorithm, and the robot can perform safety inspection tasks; for a humanoid robot, the robot algorithm can be a vision-language-motion model and a whole-body control algorithm, and the robot can perform material handling or assembly assistance tasks.

[0055] In addition, it should be noted that the simulation environment configuration used in the research, the physical parameters within the limitation range of control commands (described below), etc., can all be set according to actual needs for different types of robots.

[0056] In step S120, research can be performed iteratively based on multiple research paradigms to obtain the target algorithm corresponding to the research task.

[0057] Here, in the process of iteratively conducting research, when conducting research based on at least one of multiple research paradigms, research results obtained based on other research paradigms can be reused.

[0058] Specifically, the aforementioned research can be implemented using large models such as large language models and / or visual language models. For example, using large models, research can be conducted iteratively based on research tasks related to robotic algorithms and multiple research paradigms. For instance, multiple iterative studies can be performed, in which a complete study can be conducted for the research task in each iteration. This may include data acquisition (also known as "data sampling"), knowledge retrieval, hypothesis generation, generation of algorithm experimental packages, execution of research based on the algorithm experimental packages, evaluation of research results, and accumulation of the research findings.

[0059] As an example, research results can refer to data generated through research on robotic algorithms based on research paradigms, and may include, but are not limited to, at least one of the following: robotic algorithms; evaluation results of robotic algorithms; research experience on robotic algorithms, etc.

[0060] Specifically, in the process of iterative research, the robot algorithm obtained in each iteration can be an intermediate algorithm (e.g., an intermediate model). Here, the intermediate algorithm can be used in the next study (e.g., based on the intermediate algorithm, continue to train or optimize to obtain the algorithm for the next study), or it can not be used in the next study. For example, the intermediate algorithm can also not be input into the next study, but only the evaluation results, research experience, etc. can be input.

[0061] Evaluation results for robot algorithms can be evaluations of the performance or effectiveness of the robot algorithms obtained in this iteration of research. For example, in the model training scenario, a large model can use a test set to evaluate the model obtained in this iteration of research and obtain evaluation results. For example, the evaluation results can include indicators such as inference latency and task success rate.

[0062] Research experience regarding robot algorithms can be obtained by analyzing the above evaluation results. For example, a large model can be used to analyze the experience of this research based on the currently collected data, the current research paradigm, and the evaluation results of the algorithm. This experience may include, but is not limited to, at least one of the following: the reasons why the current algorithm did not achieve the task objective; the research direction for the next study, etc.

[0063] The above research results can be saved and accumulated during iterative research. Large models can combine the research results of all the studies that have been carried out to carry out subsequent studies, such as considering the accumulated evaluation results and drawing on previous research experience.

[0064] As an example, such as Figure 2 As shown, in each iteration, the research can be performed in the following manner: In step S210, the research results of this iteration can be obtained based on the research paradigm of this iteration and the accumulated set of research results.

[0065] Here, in the initial iteration, the accumulated research results set can be a pre-defined set, such as an empty set, or a set of historical research results. In subsequent iterations, this set can accumulate research results from previous iterations, and all research results for the same research task will be added to this set, regardless of the research paradigm from which they originate. For example, the accumulated research results set can include, but is not limited to, an experimental history repository and a domain knowledge base. The experimental history repository can store structured experimental nodes generated in each iteration, with each node including the research paradigm used in the experiment, the algorithm experimental package, the evaluation results, and the research experience. Each piece of knowledge in the domain knowledge base is accompanied by a domain tag and a confidence score. During retrieval, the data is filtered by the current task's domain tag to prevent cross-domain knowledge contamination, and an expiration elimination mechanism based on confidence score is implemented.

[0066] Furthermore, the research paradigm for the first iteration can be arbitrarily specified or selected based on the task characteristics of the research task, while the research paradigm for subsequent iterations can be determined according to the paradigm switching conditions described below.

[0067] As an example, step S210 may include: using a large language model (LLM), generating an algorithm experiment package based on the research paradigm of this iteration and the accumulated set of research results; and executing robot algorithm experiments based on the algorithm experiment package to obtain the research results of this iteration.

[0068] Here, the algorithm experiment package may include data for performing robot algorithm experiments. For example, an algorithm experiment package can refer to a structured product generated by a large language model based on research hypotheses, which can be directly executed once in a sandbox for a robot algorithm experiment. The algorithm experiment package may include, but is not limited to, model code, training scripts, simulation environment configuration, evaluation protocol, and safety parameters. For example, taking the hypothesis of "reducing vision-language-action (VLA) inference latency" as an example, an algorithm experiment package may include: (1) model code that implements lightweight cross-attention alignment; (2) training scripts that specify the learning rate, batch size, and number of training steps; (3) MuJoCo simulation configurations that specify the object, lighting, and camera pose; (4) evaluation protocols that use inference latency and task success rate as indicators; and (5) safety parameters such as maximum joint velocity and torque. Furthermore, as an example, for the aforementioned sandbox used to perform experiments, the sandbox execution infrastructure can support cross-node task scheduling, priority queuing, and breakpoint recovery after experiment interruption for multi-machine, multi-GPU clusters, and can also support a data flywheel that mixes real machine data with simulation data for retraining after quality screening.

[0069] As an example, in each iteration, before generating the algorithm experiment package, a predetermined number of context experiment nodes can be selected from the accumulated research results set (e.g., a shared experimental history library) according to the strategy of the current research paradigm (e.g., sampling strategy and / or expansion strategy). Based on these context experiment nodes and the retrieved knowledge, constrained prompts can be generated, and then LLM can generate hypotheses based on the prompts and transform them into the algorithm experiment package.

[0070] For example, taking the evolutionary loop research paradigm, the evolutionary loop can employ pluggable sampling strategies, including UCB1 upper confidence bound sampling, Island and MAP-Elites joint sampling, greedy sampling, and random sampling, which can be switched according to configuration via a factory pattern. UCB1 (Upper Confidence Bound 1) can select context experimental nodes by calculating the upper confidence bound on historical nodes. Taking the best-priority tree search as an example, it can follow a four-stage process of "draft-improvement-ablation-finalization," expanding the expandable node with the highest score at each stage to select the context experimental node. Taking the minimalist fast loop as an example, it can follow a process of "single-file editing-fixed-time-budget execution-Keep / Revert binary decision" to select the context experimental node.

[0071] As an example, in step S210, LLM can be used to parse the task description of the current research task, and based on the sampling strategy of the current research paradigm in this iteration, multiple contextual experimental nodes can be selected. Then, Top-K knowledge entries can be filtered by domain label from the accumulated research results set. As an example, the P99 (99th percentile) latency consumed in retrieving knowledge can be less than 500ms; the specific value can depend on the size of the knowledge base and the hardware configuration.

[0072] For example, constrained cue words can be configured, LLM can be used to generate structured hypotheses for the current research task, and LLM can be used as an experiment designer to transform structured hypotheses into executable algorithmic experiment packages.

[0073] Once you have the algorithm experiment package, you can organize robot algorithm experiments according to the requirements of each experiment node in the package, such as training AI models according to the requirements in the package.

[0074] For example, taking the "lightweight cross-attention alignment" algorithm experimental package mentioned above as an example, during execution, the algorithm experimental package can first undergo a static code security check. If the check passes, the VLA model can be trained in a runtime container isolated from the outside according to the training script in the algorithm experimental package. Furthermore, it can run according to the evaluation protocol in the scenario specified in the simulation configuration of the algorithm experimental package, collecting inference latency and task success rate data. The multimodal evaluator can output a composite score based on the collected data, and the statistical verification module can determine whether the improvement is significant, obtaining multidimensional evaluation results, accumulated experimental nodes, and the experience of this iteration (e.g., "lightweight cross-attention can reduce latency"). Here, the multidimensional evaluation results, accumulated experimental nodes, and the experience of this iteration can be used as the research results of this iteration. Then, the data from this executed experiment can be written into a shared experimental history library as structured experimental nodes, and the experience of "lightweight cross-attention can reduce latency" can also be written into a shared domain knowledge base. Furthermore, the aforementioned multidimensional evaluation can be performed in parallel by a numerical indicator calculator, a visual language model visual analyzer, a security evaluator, and a statistical validator, and aggregated by a score aggregator using one of the following: threshold-filtered weighted summation, Pareto front, or lexicographical order.

[0075] In step S220, in response to the current research progress meeting the paradigm switching condition, the research results of this iteration can be added to the accumulated research results set, another research paradigm among multiple research paradigms can be used as the research paradigm for the next iteration, and the next iteration can be executed.

[0076] As an example, paradigm shift conditions may include at least one of the following: based on the accumulated set of research results and the research results of this iteration, a dominant paradigm suitable for the current research progress is identified among multiple research paradigms; the budget allocated to the research paradigm of this iteration has been exhausted.

[0077] Specifically, in one example, a dominant paradigm can be determined based on the task characteristics of the current research progress, and then switched to that dominant paradigm in the next iteration.

[0078] As an example, task features can be features that characterize the research tasks involved in the current research progress (e.g., sub-tasks or phased tasks in the entire research process). Different research tasks can correspond to different task features. The research tasks here can be set according to actual needs or planned or generated by a large language model during the research process. For example, they can be known tasks in the field of robot algorithms, such as architecture search, policy exploration, ablation analysis, depth analysis, hyperparameter iteration, fast iteration, etc.

[0079] For example, in examples of multiple research paradigms including tree search, evolutionary loop, and minimalist loop, if the current research progress indicates that the current research includes architecture search and / or strategy exploration, then evolutionary loop can be considered the dominant paradigm; if the current research progress indicates that the current research includes ablation analysis and / or deep analysis, then tree search can be considered the dominant paradigm; and if the current research progress indicates that the current research includes hyperparameter iteration and / or rapid iteration, then minimalist loop can be considered the dominant paradigm.

[0080] In another example, the switch can be performed based on the task characteristics of the current research progress and the budget size of the current research paradigm.

[0081] For example, taking multiple research paradigms including tree search, evolutionary loop, and minimalist loop as examples, if the current research progress indicates that the current research task includes architecture search and / or strategy exploration, and the budget of the currently adopted research paradigm is greater than a first threshold, then evolutionary loop can be considered the dominant paradigm. If the current research progress indicates that the current research task includes hyperparameter iteration and / or rapid iteration, and the budget of the currently adopted research paradigm is less than a second threshold, then minimalist loop can be considered the dominant paradigm. Furthermore, if the current research progress indicates that the current research task includes ablation analysis and / or deep analysis, then tree search can be considered the dominant paradigm; this determination can be independent of the size of the research paradigm's budget.

[0082] Through the above methods, various research paradigms, such as evolutionary loops, best-priority tree search, and minimalist fast loops, can be unified into a single research orchestration process. Each paradigm can share the same experimental history repository, domain knowledge base, and sandbox execution infrastructure. During the research process, paradigms are automatically selected or switched at runtime according to the task characteristics of the research progress, and accumulated experimental nodes and empirical rules can be reused across paradigms. For example, effective strategies discovered in evolutionary loops can be directly reused in the tree search improvement stage, thereby eliminating information barriers between paradigms. A single task can switch paradigms on demand during execution without migrating data or switching systems.

[0083] Furthermore, as an example, the steps described above, which involve using another research paradigm among the multiple research paradigms as the research paradigm for the next iteration and executing the next iteration, may include: replacing the strategy object of the current iteration's research paradigm with the strategy object of the other research paradigm, while keeping the storage handle of the accumulated research results set unchanged, so that the other research paradigm executes the next iteration based on the research results set accumulated before the switch.

[0084] Specifically, during paradigm shifts, only the sampling strategy and / or extension strategy objects can be replaced, while the storage handles pointing to the shared experimental history library and the shared domain knowledge library (i.e., the same references used by each research paradigm to access the above-mentioned shared libraries, hereinafter referred to as library handles) remain unchanged. The shared libraries are not cleared, and the new paradigm can continue its research from the accumulated nodes and knowledge. This is also the essential difference between the unified orchestration proposed in the embodiments of this disclosure and the parallel operation of multiple systems.

[0085] Furthermore, in step S220, for example, an evaluator can be used to output a composite score of the research results of this iteration, and structured experimental nodes can be written into a shared experimental history database, and experience can be extracted and written into a shared domain knowledge base. The above writing operations are visible to all research paradigms and are the physical carriers for cross-paradigm knowledge transfer.

[0086] Furthermore, in embodiments of this disclosure, all research paradigms can be assembled as pluggable strategies via a factory pattern, bound to the same ExperimentDB / CognitionStore / SandboxExecutor handle.

[0087] Return to reference Figure 2 Alternatively or additionally, in step S230, in response to the current research progress not meeting the paradigm switching condition and not meeting the real machine migration condition, the research results of this iteration can be added to the accumulated research results set, the research paradigm of this iteration can be used as the research paradigm of the next iteration, and the next iteration can be executed.

[0088] In this step, if the paradigm shift condition is not met, it can be determined whether the real-device migration condition is met. If the real-device migration condition is also not met, the next iteration of research can continue based on the current research paradigm. Here, the real-device migration condition can be set according to actual needs.

[0089] Return to reference Figure 2 Alternatively or additionally, in step S240, in response to the current research progress not meeting the paradigm switching condition but meeting the real machine migration condition, the target algorithm can be obtained based on the research results of this iteration.

[0090] In this step, if the current research progress meets the conditions for real-device migration, the research results of this iteration can be used as the output of the target algorithm.

[0091] The above approach can solve the problems of existing autonomous research systems that adhere to a single paradigm, lack infrastructure sharing, and cannot reuse the same experimental accumulation between paradigms, forcing researchers to switch between systems. It eliminates information barriers between paradigms, can automatically select and switch at runtime according to task characteristics, and experimental nodes and empirical rules produced by one paradigm can be directly reused by another paradigm. For the same task, paradigms can be switched on demand during execution without data migration.

[0092] Figure 3 An example process for orchestrating multi-paradigm unification studies according to embodiments of this disclosure is illustrated.

[0093] like Figure 3 As shown, in step S310, the current research paradigm for the first iteration can be selected or specified according to the task characteristics.

[0094] In step S320, a single-step research loop can be executed. For example, based on the current research paradigm, a process of sampling, retrieval, hypothesis generation, experimental package generation, experiment execution, evaluation, and research result accumulation can be performed. In this step, read and write operations can be performed on the infrastructure shared by each paradigm. For example, resources for generating experimental packages can be obtained from the experimental history repository and domain knowledge base, and experiments can be executed using a sandbox executor. Here, each paradigm can bind to the same library handle.

[0095] In step S330, it can be determined whether the current research progress meets the paradigm shift conditions. In response to meeting the paradigm shift conditions, in step S340, the paradigm can be switched. During the switching process, only the sampling strategy and / or extension strategy objects are replaced, the library handle remains unchanged, and the accumulated research results are reused.

[0096] In response to the failure to meet the paradigm shift conditions, such as the failure to identify a dominant paradigm and / or the failure to exhaust the budget of the current paradigm, in step S350, it can be further determined whether the current research results meet the real machine migration conditions. In response to meeting the real machine migration conditions, the research can be terminated and the target algorithm can be output. In response to the failure to meet the real machine migration conditions, the current paradigm can be maintained and the execution step S320 can be returned.

[0097] Return to reference Figure 1 In step S130, the robot can be controlled to perform tasks based on the target algorithm.

[0098] Here, the tasks performed by the robot can be set according to actual needs.

[0099] In one example, the target algorithm can be directly migrated (or deployed) to the robot, and the robot can be controlled to perform tasks.

[0100] In another example, the target algorithm can be optimized at least once to obtain an optimized target algorithm; based on the optimized target algorithm, the robot can be controlled to perform tasks. Specifically, the optimized algorithm can be migrated (or deployed) to the robot, and then the robot can be controlled to perform tasks based on this algorithm.

[0101] Here, each algorithm optimization may include: according to multiple preset speed levels, in order of increasing speed, controlling the robot to perform tasks based on the algorithm to be optimized at each speed level to obtain real machine data. In the first optimization, the algorithm to be optimized is the target algorithm; based on the real machine data, the target algorithm is optimized to obtain the algorithm optimized in this optimization; in response to the algorithm optimized in this optimization meeting the preset output conditions, the algorithm optimized in this optimization is used as the optimized target algorithm; in response to the algorithm optimized in this optimization not meeting the above preset output conditions, the algorithm optimized in this optimization is used as the algorithm to be optimized in the next optimization and the next optimization is performed.

[0102] As an example, the above algorithm optimization can be implemented in a multi-stage Sim2Real bridging process, which may include five stages executed sequentially: simulation domain randomization, gap analysis, security checkpoint verification, finite real machine verification, and feedback loop. Each algorithm optimization includes at least finite real machine verification and feedback loop.

[0103] Specifically, such as Figure 4 As shown, in step S410, stage one, simulation domain randomization, can be performed. Simulation domain randomization can refer to progressively expanding the randomization range of visual and / or physical and / or dynamic parameters, batch evaluating each set of parameters generated by the target algorithm. Simulation domain randomization can be implemented using methods such as Automatic Domain Randomization (ADR). In step S420, it can be determined whether the success rate under the maximum randomization range is greater than or equal to a threshold. If it is less than the threshold, step S410 can be returned to continue expanding the parameter range or retraining; if it is greater than or equal to the threshold, step S430 can be performed to enter stage two.

[0104] In step S430, stage two, gap analysis, can be performed. Gap analysis can be used to compare the gap between the simulated rendered frame and the real machine image. For example, a visual language model can be used to automatically analyze visual differences. Here, visual differences may include, but are not limited to, differences in at least one of texture, lighting, and background complexity. Based on the visual differences, physical parameters can be identified, generating a quantified simulation-to-reality gap (Sim-Real Gap, or "Sim2Real Gap") vector and parameter correction suggestions. Unlike schemes that only prompt for in-word parameter tuning, in this stage two of the embodiments of this disclosure, a structured gap that can be returned can be generated.

[0105] Here, the generated gap vector and parameter correction suggestions can be used as one of the inputs to the safety checkpoint verification of Phase Three performed in step S440. Furthermore, as an example, in step S480 described below, during the execution of the feedback loop of Phase Five, the gap vector and parameter correction suggestions generated in step S430 can also be used as input to step S410, along with the actual machine data, to write back and update the randomization range of the simulation domain in Phase One, for identifying and correcting the simulation physical parameters to reduce the gap.

[0106] Furthermore, as an example, a visual language model (such as a visual language model visual analyzer) can be used to automatically analyze simulated rendering frames and real device recordings. This can include, but is not limited to, extracting failed trajectories (episodes), analyzing keyframes to determine the reasons for failure, evaluating trajectory smoothness, and comparing visual differences between simulated and real device frames to obtain the aforementioned structured differences. The number of keyframes extracted for each trajectory can be less than or equal to 10, and the specific number can be adjusted according to the trajectory length and analysis accuracy requirements.

[0107] In step S440, Phase Three: Safety Checkpoint Verification can be executed. In safety checkpoint verification, at least one of the following operations can be performed: automatic safety pre-checks, such as reviewing dangerous calls to policy code; simulated stress testing, such as batch verification under extreme conditions; manual approval, such as automatically submitting safety reports to approvers without any system parameters or APIs that can be skipped; and real-machine safety parameter injection, such as writing safety parameters like maximum joint speed, torque, end effector force, and workspace boundaries into the control pipeline. In step S450, it can be determined whether manual approval has passed and whether all stress tests have passed. If manual approval has passed and all stress tests have passed, step S460 can be executed to proceed to Phase Four; if manual approval has failed and / or all stress tests have failed, the process can return to step S440 for rectification, and the approval decision can be permanently recorded in an append-only audit log.

[0108] In step S460, Phase Four: Limited Real Machine Verification can be performed. Limited real machine verification can be performed by progressively increasing the robot's movement speed (e.g., from 30% to 60% to 100% of the rated speed), monitoring torque and / or position over-limits and collisions in real time during operation, and immediately stopping upon any trigger. It can also automatically collect robot parameters such as RGB, depth, joints, forces, and end-effector pose, and automatically determine the execution status of the task based on preset thresholds. Here, the execution status may include, for example, success (Pass), failure (Fail), and need more data (NeedMoreData).

[0109] As an example, in response to the robot failing to execute a task at any speed level, the algorithm to be optimized can be adjusted, and the robot can be controlled to execute the unsuccessfully executed task based on the adjusted algorithm until all tasks at the current speed level are successfully executed, and then switch to the next speed level; in response to the robot successfully executing a task at the highest speed level, the data generated during the execution of tasks at all speed levels can be used as real machine data.

[0110] Here, the failure of a robot task can be determined by: monitoring the robot's operating parameters in real time during the task execution; and triggering an emergency stop of the robot in response to the monitoring that the operating parameters meet preset constraints, thus confirming that the current task has not been successfully executed. The operating parameters may include at least one of torque, position, and collision risk.

[0111] As an example, preset constraints can be set according to actual needs. For example, limits can be set for torque, position, and collision risk. If the torque, position, and collision risk exceed the corresponding limits, the task can be considered unsuccessful.

[0112] For example, such as Figure 4 As shown, in step S470, it can be determined whether the task was successfully executed at the highest speed setting. If the task was not successfully executed, more data can be added, the target algorithm can be adjusted, and the execution can return to step S460. If the task was successfully executed, the target algorithm can be transferred to the robot, and step S480 can be executed to enter stage five.

[0113] In step S480, stage five: feedback loop closure, can be executed. During the feedback loop closure process, it can be determined whether the optimized algorithm meets preset output conditions. These preset output conditions may include, for example, whether the current number of optimization iterations has reached a preset number; and / or whether the current optimized algorithm has reached a preset target. If the above conditions are met, the target algorithm of the current optimization can be transferred to the robot to complete the transfer and execute subsequent control; if the above conditions are not met, the next optimization can be executed.

[0114] As an example, if the above conditions are not met, the real machine dynamics data from the current optimization process can be used as input for the next optimization, for example... Figure 4 As shown, the real machine dynamics data can be fed back to step S410. Here, the real machine dynamics data may include, but is not limited to, at least one of the following: fine-tuning simulation physical parameters, such as friction, mass, and damping; real machine observation distribution, such as feedback to update the randomization range of the visual domain; failure cases, such as extracting empirical rules and writing them into a shared domain knowledge base; and successful configurations, such as accumulating them into best practices. The fed-back data can be used for subsequent verification and closed-loop calibration of the target algorithm.

[0115] Using the above methods, real device permissions can be gradually released under controlled conditions, and Sim2Real Gap can be continuously reduced based on feedback from the real device.

[0116] The foregoing describes the multi-paradigm iterative execution study and the process of transferring the algorithm to a robot according to embodiments of the present disclosure. As an example, security checks can be set separately in the above iterative study and algorithm transfer process to ensure the reliability of the robot's task execution.

[0117] As an example, the steps described above for executing robot algorithm experiments based on the algorithm experiment package to obtain the research results of this iteration may include: performing security verification on the generated algorithm experiment package; intercepting the algorithm experiment package in response to it failing the security verification; and / or, in response to the algorithm experiment package passing the security verification, executing robot algorithm experiments in a runtime container based on the algorithm experiment package to obtain the research results of this iteration.

[0118] Here, the algorithm running inside the runtime container can be isolated from the outside of the runtime container.

[0119] Specifically, security checks can be performed on algorithm experiment packages to achieve static code security control before execution, intercepting dangerous calls. Experiment packages that pass security checks can be executed in isolated containers. As an example, security checks may include, but are not limited to, Abstract Syntax Tree (AST) scanning and injection protection. Here, AST scanning can refer to parsing the code's AST, matching dynamic attribute access and / or string concatenation call patterns, and if a match is found, code interception can be performed and written to the audit log.

[0120] For example, you can apply a blacklist of dangerous modules (such as disabling os / subprocess / socket, etc.), disable dangerous built-in functions (such as eval / exec / compile / __import__), perform deep scanning based on abstract syntax trees (such as detecting string concatenation or dynamic access to attributes bypass, such as getattr(os,'system')), and protect against prompt injection to the code generated by LLM. If the verification passes, it will enter the container for execution; if the verification fails, it will be intercepted and recorded in the audit.

[0121] Furthermore, as an example, the aforementioned security verification may also include dependency whitelist verification, whereby the dependency whitelist limits the allowed imported dependency libraries in the algorithm experimental package, such as torch, numpy, mujoco, gymnasium, etc. For instance, when performing security verification on the generated algorithm experimental package, the dependency libraries imported by the package can be parsed. If a dependency library outside the whitelist is found, the algorithm experimental package is determined to have failed the security verification and is blocked, and written to the audit log; if all imported dependency libraries belong to the whitelist, the remaining verification items are executed. The security verification is enforced on all algorithm experimental packages, and there is no bypass to skip the security verification.

[0122] Furthermore, in embodiments of this disclosure, each experiment can be run within an independent runtime container. Here, each runtime container can be network isolated from external sources (e.g., other runtime containers or other running programs), non-rooted, whitelisted for seccomp system calls, and cgroup-limited to CPU / memory / video memory / disk.

[0123] In the embodiments of this disclosure, in addition to performing the above-mentioned security verification process on the algorithm experimental package, physical hard constraints on the actual machine can also be performed after the algorithm is deployed on the robot to further ensure the safety of the robot's actions.

[0124] As an example, the steps described above for controlling a robot to perform a task based on a target algorithm may include: generating control instructions for controlling the robot to perform a task based on the target algorithm; triggering an emergency stop of the robot in response to the control instructions exceeding a preset limit; and controlling the robot to perform a task based on the control instructions in response to the control instructions not exceeding the limit.

[0125] Here, the limitation range can be used to restrict the instructions in the control commands that are related to the robot's physical parameters.

[0126] Furthermore, the target algorithm here can be an initial target algorithm or an optimized target algorithm obtained by performing at least one algorithm optimization as described above. Specifically, algorithm optimization and limit truncation can be implemented in combination. As an example, the process of controlling the robot to perform a task based on the optimized target algorithm can include: generating control commands for controlling the robot to perform the task based on the optimized target algorithm; triggering an emergency stop of the robot in response to the control commands exceeding a preset limit range; and controlling the robot to perform the task based on the control commands in response to the control commands not exceeding the limit range.

[0127] Specifically, before the control command is issued, the policy output actions in the control command can be forcibly truncated one by one. For example, the action action_exec sent to the robot can be represented as: action_exec = clamp(action_brain, safety_params), where clamp means truncation, action_brain means policy output action in the control command, and safety_params means the limit range.

[0128] As an example, the physical parameters of the robot described above may include, but are not limited to: joint positions (e.g., URDF-limited), joint velocities, joint torques, end effector forces, and workspace boundaries (e.g., Cartesian workspace boundaries). In response to any physical parameter exceeding its limits, such as at least one of the following: joint velocity, joint torque, or end effector force exceeding a pre-configured limit; violation of workspace boundary limits; collision; or communication timeout, a hardware emergency stop can be triggered. Here, the emergency stop is executed at the hardware level and cannot be disabled by software. Furthermore, as an example, during communication with the robot, digital signatures and / or two-way authentication can be applied to control commands sent to the robot. If a control command fails to reach a threshold time, a communication timeout can be considered to have occurred, and the robot can be safely stopped.

[0129] The above approach enables asymmetric depth from static code control to physical hard constraints, ensuring that even if a large model generates out-of-bounds instructions, the physical layer will forcibly truncate and prevent execution before the control instructions are issued, and the physical layer architecture cannot be bypassed by the upper layer.

[0130] Figure 5 An example architecture of a two-layer asymmetric security control according to an exemplary embodiment of the present disclosure is shown.

[0131] like Figure 5As shown, to ensure the safety of robot actions, static code control can be performed on the algorithm experimental package before execution at the upper layer, and unsafe code can be intercepted. Specifically, in step S511, an experimental package (e.g., experimental code) can be generated using LLM. In step S512, static code control can be performed on the experimental package, such as performing AST scanning, filtering blacklists and whitelists, and injection protection. In step S513, it can be determined whether the verification passes. In response to failure, in step S514, the algorithm experimental package can be intercepted and written to the audit log; in response to success, in step S515, the algorithm experimental package can be executed in isolation within a container.

[0132] In addition, such as Figure 5 As shown, to ensure the safety of robot movements, a physical hard constraint can be implemented at the lower layer for the target algorithm deployed to the robot. Specifically, in step S521, control commands can be generated based on the target algorithm. In step S522, based on the preset physical hard constraint layer, it can be determined whether the control commands exceed the preset limit range. In response to any parameter (such as joint, speed, torque, workspace, etc.) exceeding the corresponding limit range, in step S523, the robot's hardware emergency stop is triggered. In response to all parameters not exceeding the corresponding limit range, in step S524, the motor is driven based on the control commands to control the robot to perform the task.

[0133] Using the above method, code generation for LLM can achieve two-layer asymmetric safety control of physical robots. The upper layer is static code control and container isolation before execution, while the lower layer is a physical hard constraint layer that acts on the real machine and forcibly truncates before the control command is issued, and a hardware emergency stop that cannot be disabled by software. For the lower layer, no upper-layer software or large model command can bypass the judgment of the physical hard constraint layer to bypass the control command. In this way, even if the large model generates an out-of-bounds command in the algorithm, the physical layer will forcibly truncate it before the control command is issued and will not execute it.

[0134] Figure 6 This is a schematic flowchart illustrating an application example of a robot control method according to an exemplary embodiment of the present disclosure. Figure 7 This is a schematic diagram illustrating an example architecture for implementing a control method for a robot according to an exemplary embodiment of the present disclosure.

[0135] like Figure 6 As shown, in step S601, a task can be received, and the current research paradigm can be selected based on the task characteristics. In step S602, based on the current research paradigm, hypotheses can be generated using constrained LLM sampling context and retrieved knowledge, and an algorithm experimental package can be generated based on the hypotheses. For example, as... Figure 7As shown, the multi-paradigm can be uniformly orchestrated and scheduled by the research orchestrator 701 to determine the current research paradigm. Based on the current research paradigm, the shared knowledge and data layer 702 can be sampled, retrieved, and accumulated to obtain the algorithm experimental package.

[0136] In step S603, pre-execution static code security control can be performed. After passing the verification, the code can be executed within the isolated container. For example... Figure 7 As shown, static code management of experimental packages can be performed in the experimental execution layer 703, and experimental packages can be executed in a containerized sandbox.

[0137] In step S604, the execution results can be evaluated in multiple dimensions, and the structured records can be written to the shared experimental history database, while the extracted experience can be written to the shared domain knowledge base. For example... Figure 7 As shown, the execution results can be evaluated using the multimodal evaluator 704, including dimensions such as numerical, VLM, security, and statistics, and the research results can be written back to the shared knowledge and data layer.

[0138] In step S605, it can be determined whether the paradigm shift condition is met. If the paradigm shift condition is met, in step S606, the research can be switched to another paradigm to continue execution, and the results accumulated in the same experimental history repository / domain knowledge base / sandbox infrastructure can be reused after the switch. If the paradigm shift condition is not met, in step S607, it can be determined whether migration to a live machine is necessary.

[0139] If migration to the real machine is not required, the process can return to step S602; if migration to the real machine is required, in step S608, simulation domain randomization can be performed, and a "simulation-real machine" gap analysis can be performed.

[0140] In step S609, manual approval can be forced, and the outcome of the manual approval can be determined. If the manual approval fails, the process can return to step S608; if the manual approval passes, in step S610, a progressively accelerated limited real-device verification can be performed. For example, as... Figure 7 As shown, in response to triggered migration, simulation domain randomization and “simulation-real machine” gap analysis can be performed in the simulation and hardware abstraction layer 705.

[0141] In step S611, a physical hard limit truncation can be performed, and the real machine data can be fed back to the calibration simulation process. For example... Figure 7 As shown, the security management system 706 can be used to implement code-level control and physical hard restrictions. Ultimately, the algorithm can be deployed to a real robot, and fax machine data can be fed back to calibrate the simulation.

[0142] Several application examples of the robot control method according to embodiments of the present disclosure will be described below.

[0143] In one example, the system receives a research task input by the user: "Explore more efficient visual-language alignment methods to reduce inference latency in visual-language-action models and maintain the success rate of grasping tasks." It also receives a total budget allocated to this research task, such as 60 iterations. After parsing the research task, three research paradigms can be obtained: evolutionary loop, best-priority tree search, and minimalist fast loop. The total budget is then divided according to a preset ratio, allocating budgets to each paradigm. For example, 30 iterations are allocated to the evolutionary loop, 20 iterations to the best-priority tree search, and 10 iterations to the minimalist fast loop.

[0144] In the first iteration, considering that the research task's characteristics include architecture search and the budget allocated to the evolutionary loop is greater than a first threshold, the evolutionary loop can be automatically recommended as the research paradigm for the first iteration.

[0145] Specifically, contextual experimental nodes can be selected from a shared experimental history database using an evolutionary sampling strategy (no experimental nodes are present in the database during the first iteration, and the sampling strategy is random initialization). Top-K knowledge entries can be retrieved from a shared domain knowledge base based on domain tags to generate constrained prompts. A structured research hypothesis is generated using a large language model: "Replace the original cross-attention with lightweight cross-attention alignment." The large language model, acting as an experiment designer, transforms this research hypothesis into an algorithmic experimental package. This package can include model code for implementing lightweight cross-attention alignment, training scripts specifying the learning rate and number of training steps, simulation environment configurations specifying objects and lighting, evaluation protocols based on inference latency and task success rate, and safety parameters such as maximum joint speed and torque. The algorithm experimental package undergoes security verification. After successful verification, the robot algorithm experiments specified in the package are executed in a runtime container isolated from the external network. The execution results are evaluated in parallel by a numerical metric calculator, a visual language model visual analyzer, a security evaluator, and a statistical verifier. A score aggregator then aggregates the results into a composite score. For example, if the evaluation shows reduced inference latency while maintaining a roughly equal success rate for grasping tasks, the statistical verifier determines that the improvement is significant. The structured experimental nodes corresponding to this experiment are written to a shared experimental history repository, and the empirical rule "lightweight cross-attention can reduce inference latency" is written to a shared domain knowledge base.

[0146] In several iterations, no suitable paradigm for the current research progress was identified, and the evolutionary cycle budget was not exhausted, thus failing to meet the paradigm shift condition. Furthermore, the inference latency in the research results of each iteration was not lower than the preset latency threshold, also failing to meet the real-machine transfer condition. Here, the research results of each iteration were added to the accumulated research results set.

[0147] In the aforementioned iterations, during the second iteration, the generated algorithm experiment package contained a code snippet that dynamically accessed system commands through string concatenation (getattr(os,'system')). This caused the algorithm experiment package to fail the security check, resulting in its interception, and the interception event was written to the audit log. Subsequently, based on the reason for the interception, the algorithm experiment package was regenerated, passed the re-verification, and was then executed in the runtime container to continue the iteration.

[0148] During the 13th iteration, after several iterations of extensive exploration in the architecture space through the evolutionary loop, several candidate attention alignment architectures with similar composite scores emerged in the accumulated research results set. The task characteristics corresponding to the current research progress changed from architecture search to ablation analysis. In response, the optimal priority tree search was identified as the advantageous paradigm suitable for the current research progress, satisfying the paradigm switching condition. The strategy object of the evolutionary loop was replaced by the strategy object of the optimal priority tree search. In the next iteration, the optimal priority tree search can directly use the experimental nodes written into the shared experimental history library in the evolutionary loop stage as scalable nodes, without having to re-execute the experiments already completed in the evolutionary loop stage or migrate data.

[0149] After the 33rd iteration, the best priority tree search has completed 20 iterations and its budget has been exhausted, satisfying the paradigm shift condition. It then switches to a minimal fast loop, which rapidly fine-tunes its hyperparameters based on the finalized architecture accumulated in the shared library.

[0150] After the 38th iteration, the inference latency in the iterative research results was lower than the preset latency threshold, and the success rate of the grasping task in the simulation environment with the maximum randomization range was not lower than the preset success rate threshold, thus meeting the conditions for real-device transfer. Therefore, based on the research results of this iteration, the robot algorithm therein was selected as the target algorithm output corresponding to this research task.

[0151] In the above example, both the best priority tree search and the minimal fast loop can reuse the experimental nodes and empirical rules obtained from the evolutionary loop and written into the shared library when performing research. Furthermore, since the storage handle remains unchanged and the shared library is not cleared when switching paradigms, the research paradigm after the switch does not need to repeat the experiments completed before the switch. Therefore, the computational resources and execution time occupied by repeated experiments are reduced within the total budget of the same research task, allowing more effective iterations to be completed before the budget is exhausted, thereby improving the quality of the target algorithm and enabling more precise control of the robot to perform tasks.

[0152] In the embodiments of this disclosure, algorithm optimization was also performed on the target algorithm obtained from the above application example to obtain an optimized target algorithm, and an experiment was conducted to control the robot to perform a grasping task based on the optimized algorithm.

[0153] Specifically, in the first optimization, the randomization range of visual, physical, and dynamic parameters was progressively expanded through simulation domain randomization, and the target algorithm was evaluated in batches under each set of randomization parameters. The success rate of the grasping task under the maximum randomization range decreased as the randomization range increased, but it was still not lower than the preset pass threshold. To address this, a visual language model was used to compare the simulation rendering frames and real device images to identify visual differences such as "the real device background is more complex, there are side lighting shadows, and the reflective properties of the grasped object are different." Combined with physical parameter identification, a quantified vector of the gap between simulation and reality and parameter correction suggestions were generated.

[0154] Then, the algorithm's code was verified for safety, and no dangerous calls were found. Furthermore, simulation stress tests were conducted to verify the execution of the task under extreme conditions. The simulation results confirmed that there were no collisions and that the peak torque was below the limit.

[0155] Subsequently, the current algorithm was migrated to a real device for testing. Specifically, at 30% of the rated speed, the robot was controlled to perform a grasping task based on the algorithm. It was found that in the grasping task of reflective objects, the robot's end effector force met the preset limit conditions in real time monitoring, which triggered the robot to stop suddenly and determined that the grasping task was not successfully executed.

[0156] In this case, based on the parameter correction suggestions generated above, the algorithm is adjusted. Specifically, the processing of highlight areas in the visual feature extraction part of the algorithm is adjusted, and the robot is re-controlled to execute the unsuccessful grasping task based on the adjusted algorithm until all grasping tasks at the 30% speed level are successfully executed. When switching to the 60% speed level and the 100% speed level (the highest speed level), all tasks are successfully executed.

[0157] After the robot successfully performed the task at 100% speed, the algorithm was optimized based on the real machine data (including RGB images, depth images, joint positions, joint torques and end-effector poses) collected during the task execution at three speed levels. The first optimized algorithm was obtained. In this optimization, the success rate of the grasping task at 100% speed was lower than the preset success rate threshold, which did not meet the preset output conditions.

[0158] Subsequently, the above optimization process was repeated for the algorithm obtained in the first optimization. In the real device test, the crawling tasks at speed levels of 30%, 60%, and 100% were all successfully executed on the first attempt, and the data generated at the three speed levels were used as the real device data for this test. Based on this real device data, the optimized algorithm obtained in this test had a success rate of not less than the preset success rate threshold at the 100% speed level, which met the preset output conditions. The optimized algorithm was then used as the target algorithm after optimization.

[0159] The algorithm, after the two optimizations mentioned above, was deployed to the robot, and the robot was controlled to perform multiple grasping tests based on the algorithm. In one grasping test, the control command issued exceeded the preset hardware limit range, resulting in the failure to issue the control command and triggering the robot's hardware emergency stop. In the other grasping tests, the control commands did not exceed the hardware limit range and normally drove the motor to control the robot to perform grasping. The robot successfully grasped and placed the target object.

[0160] During the aforementioned control process, control commands cannot bypass the preset limit range and can be intercepted before being issued, ensuring the safety and reliability of the robot in performing its tasks.

[0161] Furthermore, in the embodiments of this disclosure, the process of this method is compared with that of traditional research paradigm generation algorithms. Specifically, by adopting the method of the embodiments of this disclosure, compared with traditional manual or single research paradigm schemes, improvements can be achieved in terms of single-step research iteration time, data migration cost for paradigm switching, Sim2Real migration cycle, real device migration security, and the interception level of out-of-bounds code. Table 1 below shows exemplary results of comparative experiments conducted on this method and traditional manual or single research paradigms.

[0162] Table 1

[0163] As can be seen from the table above, compared with traditional manual or single-paradigm solutions, the method of the embodiments of this disclosure can achieve improvements in terms of single-step research iteration time, data migration cost of paradigm switching, Sim2Real migration cycle, real-device migration security, and interception level of out-of-bounds code.

[0164] In a second aspect of exemplary embodiments of this disclosure, a control device for a robot is provided, such as... Figure 8 As shown, the control device 800 may include an acquisition unit 810, a research unit 820, and a control unit 830.

[0165] The acquisition unit 810 is configured to acquire multiple research paradigms based on a research task for robot algorithms, wherein each research paradigm refers to a methodological system used to study robot algorithms.

[0166] Research unit 820 is configured to perform research iteratively based on multiple research paradigms to obtain the target algorithm corresponding to the research task.

[0167] The control unit 830 is configured to control the robot to perform tasks based on the target algorithm.

[0168] Here, during the iterative execution of research, when conducting research based on at least one of multiple research paradigms, research results obtained based on other research paradigms are reused.

[0169] As an example, in each iteration, research unit 820 is configured to perform research in the following ways: based on the research paradigm of this iteration and the accumulated research result set, obtain the research results of this iteration, wherein in the first iteration, the accumulated research result set is a preset set; in response to the current research progress meeting the paradigm switching condition, add the research results of this iteration to the accumulated research result set, take another research paradigm among multiple research paradigms as the research paradigm of the next iteration, and execute the next iteration; and / or, in response to the current research progress not meeting the paradigm switching condition and not meeting the real machine migration condition, add the research results of this iteration to the accumulated research result set, take the research paradigm of this iteration as the research paradigm of the next iteration, and execute the next iteration; and / or, in response to the current research progress not meeting the paradigm switching condition but meeting the real machine migration condition, obtain the target algorithm based on the research results of this iteration.

[0170] As an example, a paradigm shift condition includes at least one of the following: identifying a dominant paradigm among multiple research paradigms that is suitable for the current research progress; or having exhausted the budget allocated to the research paradigm for this iteration.

[0171] As an example, research unit 820 is configured to: use a large language model to generate an algorithm experiment package based on the research paradigm of this iteration and the accumulated research results set, wherein the algorithm experiment package includes data for performing robot algorithm experiments; and perform robot algorithm experiments based on the algorithm experiment package to obtain the research results of this iteration.

[0172] As an example, the research unit 820 is configured to: perform security verification on the generated algorithm experiment package; intercept the algorithm experiment package in response to the algorithm experiment package failing the security verification; and / or, in response to the algorithm experiment package passing the security verification, execute the robot algorithm experiment in the runtime container based on the algorithm experiment package to obtain the research results of this iteration, wherein the algorithm running in the runtime container is isolated from the outside of the runtime container.

[0173] As an example, research unit 820 is configured to replace the strategy object of the current research paradigm with the strategy object of another research paradigm, while keeping the storage handle of the accumulated research results set unchanged, so that the other research paradigm executes the next iteration based on the research results set accumulated before the switch.

[0174] As an example, the control unit 830 is configured to: generate control instructions for controlling the robot to perform tasks based on a target algorithm; trigger an emergency stop of the robot in response to the control instructions exceeding a preset limit range; and control the robot to perform tasks based on the control instructions in response to the control instructions not exceeding the limit range, wherein the limit range is used to restrict the instructions in the control instructions that are related to the physical parameters of the robot.

[0175] As an example, the control unit 830 is configured to: perform at least one algorithm optimization for the target algorithm to obtain an optimized target algorithm; control the robot to perform a task based on the optimized target algorithm, wherein each algorithm optimization includes: according to multiple preset speed levels, controlling the robot to perform a task based on the algorithm to be optimized at each speed level from low to high to obtain real machine data, wherein in the first optimization, the algorithm to be optimized is the target algorithm; optimize the target algorithm based on the real machine data to obtain the algorithm optimized in this instance; in response to the algorithm optimized in this instance meeting preset output conditions, use the algorithm optimized in this instance as the optimized target algorithm; in response to the algorithm optimized in this instance not meeting preset output conditions, use the algorithm optimized in this instance as the algorithm to be optimized in the next instance and perform the next optimization.

[0176] As an example, the control unit 830 is configured to: adjust the algorithm to be optimized in response to the robot failing to execute a task at any speed level, and control the robot to execute the failed task based on the adjusted algorithm until all tasks at the current speed level are successfully executed, and then switch to the next speed level; and in response to the robot successfully executing a task at the highest speed level, use the data generated during the execution of tasks at all speed levels as real machine data.

[0177] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0178] According to a third aspect of this disclosure, an electronic device is provided. Figure 9 This is a block diagram of an electronic device according to exemplary embodiments of the present disclosure. Figure 9 As shown, the electronic device 10 includes a processor 101 and a memory 102 for storing processor-executable instructions. Here, when executed by the processor, the processor causes the processor to perform the robot control method as described in the exemplary embodiments above.

[0179] As an example, electronic device 10 is not necessarily a single device, but can be a collection of any means or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. Electronic device 10 can also be part of an integrated control system or system manager, or can be configured to interface with a server locally or remotely (e.g., via wireless transmission).

[0180] In electronic device 10, processor 101 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, processor 101 may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0181] The processor 101 can execute instructions or code stored in the memory 102, which can also store data. Instructions and data can also be sent and received over a network via a network interface device, which can employ any known transmission protocol.

[0182] The memory 102 may be integrated with the processor 101, for example, by placing RAM or flash memory within an integrated circuit microprocessor. Alternatively, the memory 102 may include a separate device, such as an external disk drive, a storage array, or other storage device that can be used by any database system. The memory 102 and the processor 101 may be operatively coupled, or may communicate with each other, for example, via I / O ports, network connections, etc., enabling the processor 101 to read files stored in the memory 102.

[0183] In addition, the electronic device 10 may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, mouse, touch input device, etc.). All components of the electronic device 10 can be connected to each other via a bus and / or network.

[0184] According to a fourth aspect of this disclosure, a robot is provided, the robot including electronic devices according to this disclosure, or the robot being communicatively connected to electronic devices according to this disclosure. As an example, the robot may include an embodied intelligent robot.

[0185] According to a fifth aspect of this disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform a robot control method according to this disclosure.

[0186] Computer-readable storage media can be, for example, a memory that includes instructions. Examples of computer-readable storage media include: read-only memory (ROM), random access memory (RAM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R. LTH, BD-RE, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or ultra-fast digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the aforementioned computer-readable storage medium can run in an environment deployed in computer devices such as clients, hosts, agent devices, servers, etc. Furthermore, in one example, the computer program and any associated data, data files, and data structures are distributed across a networked computer system, such that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner through one or more processors or computers.

[0187] According to a sixth aspect of this disclosure, a computer program product is provided, including computer-executable instructions that, when executed by at least one processor, implement the robot control method according to this disclosure.

[0188] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0189] Furthermore, it should be noted that although several examples of each step have been described above with reference to the specific accompanying drawings, it should be understood that the embodiments of this disclosure are not limited to the combinations given in the examples, and the steps appearing in different drawings can be combined, which will not be exhaustive here.

[0190] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for controlling a robot, characterized in that, The control method includes: Based on the research tasks for robot algorithms, multiple research paradigms are obtained, where each research paradigm refers to a methodological system used to study robot algorithms; Based on the aforementioned multiple research paradigms, the research is performed iteratively to obtain the target algorithm corresponding to the research task; Based on the target algorithm, the robot is controlled to perform tasks. In the process of iteratively performing the research, when performing the research based on at least one of the multiple research paradigms, the research results obtained based on other research paradigms are reused.

2. The control method according to claim 1, characterized in that, In each iteration, the research is conducted in the following manner: Based on the research paradigm of this iteration and the accumulated research results set, the research results of this iteration are obtained. In the first iteration, the accumulated research results set is a preset set. In response to the current research progress meeting the paradigm shift condition, the research results of this iteration are added to the accumulated research results set, another research paradigm among the multiple research paradigms is selected as the research paradigm for the next iteration, and the next iteration is executed; and / or, If the current research progress does not meet the conditions for paradigm switching and real-device migration, the research results of this iteration are added to the accumulated research results set, the research paradigm of this iteration is used as the research paradigm for the next iteration, and the next iteration is executed; and / or, In response to the fact that the current research progress does not meet the paradigm shift condition but meets the real machine migration condition, the target algorithm is obtained based on the research results of this iteration.

3. The control method according to claim 2, characterized in that, The paradigm shift condition includes at least one of the following: It was identified that among the multiple research paradigms, there is an advantageous paradigm that is suitable for the current research progress; The budget allocated to the research paradigm for this iteration has been exhausted.

4. The control method according to claim 2, characterized in that, Based on the research paradigm of this iteration and the accumulated set of research results, the research results of this iteration are obtained, including: Using a large language model, based on the research paradigm of this iteration and the accumulated research results set, an algorithm experiment package is generated, wherein the algorithm experiment package includes data for performing robot algorithm experiments; Based on the aforementioned algorithm experimental package, robot algorithm experiments were performed, and the research results of this iteration were obtained.

5. The control method according to claim 4, characterized in that, The robot algorithm experiment, based on the aforementioned algorithm experimental package, was executed to obtain the research results of this iteration, including: Perform security checks on the generated algorithm experimental packages; In response to the algorithm experiment package failing the security check, the algorithm experiment package is intercepted; and / or, In response to the algorithm experimental package passing the security verification, a robot algorithm experiment is executed in the runtime container based on the algorithm experimental package to obtain the research results of this iteration, wherein the algorithm running in the runtime container is isolated from the outside of the runtime container.

6. The control method according to claim 2, characterized in that, The step of selecting another research paradigm from the plurality of research paradigms as the research paradigm for the next iteration and executing the next iteration includes: The strategy object of the current research paradigm is replaced by the strategy object of the other research paradigm, while keeping the storage handle of the accumulated research results set unchanged, so that the other research paradigm executes the next iteration based on the accumulated research results set before the switch.

7. The control method according to claim 1, characterized in that, The step of controlling the robot to perform tasks based on the target algorithm includes: Based on the target algorithm, control instructions are generated to control the robot to perform tasks. In response to the control command exceeding the preset limit range, the robot is triggered to stop suddenly; In response to the control command not exceeding the limit range, the robot is controlled to perform the task based on the control command. The limitation range is used to restrict the instructions in the control commands that are related to the physical parameters of the robot.

8. The control method according to claim 1, characterized in that, The step of controlling the robot to perform tasks based on the target algorithm includes: For the target algorithm, perform at least one algorithm optimization to obtain the optimized target algorithm; Based on the optimized target algorithm, the robot is controlled to perform the task. Each algorithm optimization includes: Based on multiple preset speed levels, the robot is controlled to perform tasks at each speed level from low to high, according to the algorithm to be optimized, to obtain real machine data. In the first optimization, the algorithm to be optimized is the target algorithm. Based on the real device data, the target algorithm is optimized to obtain the optimized algorithm. If the optimized algorithm meets the preset output conditions, the optimized algorithm will be used as the target algorithm after optimization. If the optimized algorithm does not meet the preset output conditions, the optimized algorithm will be used as the algorithm to be optimized in the next optimization and the next optimization will be performed.

9. The control method according to claim 8, characterized in that, The process involves controlling the robot to perform tasks based on an algorithm to be optimized, according to multiple preset speed levels, sequentially increasing the speed at each speed level, to obtain real machine data, including: In response to the robot failing to execute a task at any speed level, the algorithm to be optimized is adjusted, and the robot is controlled to execute the unsuccessfully executed task based on the adjusted algorithm until all tasks at the current speed level are successfully executed, and then the robot is switched to the next speed level. In response to the robot successfully executing the task at its highest speed setting, the data generated during the execution of the task at all speed settings will be used as the actual machine data.

10. A control device for a robot, characterized in that, The control device includes: The acquisition unit is configured to acquire multiple research paradigms based on the research task for robot algorithms, wherein each research paradigm refers to a methodological system used to study robot algorithms; The research unit is configured to iteratively perform research based on the multiple research paradigms to obtain the target algorithm corresponding to the research task; The control unit is configured to control the robot to perform tasks based on the target algorithm. In the process of iteratively performing the research, when performing the research based on at least one of the multiple research paradigms, the research results obtained based on other research paradigms are reused.

11. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store the processor's executable instructions. Wherein, when the processor executes the executable instructions, it causes the processor to perform the robot control method according to any one of claims 1 to 9.

12. A robot, characterized in that, The robot includes the electronic device according to claim 11, or the robot is communicatively connected to the electronic device according to claim 11.

13. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the robot control method according to any one of claims 1 to 9.

14. A computer program product comprising computer-executable instructions, characterized in that, When the computer-executable instructions are executed by at least one processor, they implement the robot control method according to any one of claims 1 to 9.

Images