Dynamic path correction method of artificial intelligence robot
By combining initial sequence generation and iterative optimization loops with a global attention mechanism, the problems of error accumulation and lack of global vision in the Transformer architecture are solved, achieving efficient, smooth and consistent generation of robot motion sequences.
Patent Information
- Application Number
- CN202511802366.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-17
AI Technical Summary
Existing robot motion generation methods based on the Transformer architecture suffer from problems such as error accumulation, lack of correction mechanisms, and missing global vision, resulting in disjointed robot motion and difficulty in generating consistent motion sequences.
It employs an initial sequence generation, iterative optimization loop, and global attention mechanism. Through multi-round optimization and parallel decoding, it uses a non-causal Transformer decoder to correct action sequences and combines environmental observations for dynamic path correction.
It generates temporally consistent and smooth motion sequences that can flexibly respond to environmental changes, avoid motion jitter, and improve the accuracy and consistency of motion sequences.
Smart Images
Figure CN121682771A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a dynamic path correction method for an artificial intelligence robot. Background Technology
[0002] Transformer-based models have become an important policy representation method in robot imitation learning and reinforcement learning. They typically employ encoder-decoder or decoder-only structures to generate robot action sequences in an auto-regressive manner. However, this method has two inherent drawbacks:
[0003] 1. Error Accumulation and Lack of Correction Mechanism: The autoregressive generation process is strictly unidirectional. When generating the t-th action, the model can only rely on the previously generated actions 1 to t-1 and their observations. Once there is an error or suboptimal decision in the preceding actions, this error will directly serve as the input context for the generation of subsequent actions, causing the error to propagate continuously and cannot be corrected. The entire sequence generation process lacks a global re-optimization and adjustment mechanism, which may cause the robot's behavior to deviate from the optimal path;
[0004] 2. Lack of Global Vision and Sequence Inconsistency: To ensure the mathematical correctness of autoregressive generation, the model must use a causal mask, which prevents the output at position t from considering positions t+1 to T. This means that the model cannot utilize any future state information when generating any action. This temporal constraint severely limits the model's global planning ability, making it difficult to generate highly consistent and smooth long-term action sequences, resulting in jitter and incoherence in robot motion. Summary of the Invention
[0005] The purpose of this invention is to provide a dynamic path correction method for artificial intelligence robots to solve the aforementioned background technical problems.
[0006] To achieve the above objectives, the present invention adopts the following technical solution;
[0007] A dynamic path correction method for an artificial intelligence robot, comprising:
[0008] Step 1: Initial sequence generation. The correction system receives the initial environmental observations O and quickly generates a preliminary complete action sequence A^0 = [a_1^0, a_2^0, ..., a_T^0] using an initial predictor.
[0009] Step 2: Iterative optimization loop. The system does not directly execute the action sequence A^0, but instead sends the action sequence A^0 into an iterative optimization module for N rounds of optimization, where N is greater than 1.
[0010] Step 3: Sequence execution. After N rounds of optimization, the final optimized sequence A^N is output, and the robot begins to execute the actions in the A^N sequence.
[0011] As a further description of the above technical solution:
[0012] The initial environmental observations O received by the system in step one include environmental images and robot status.
[0013] As a further description of the above technical solution:
[0014] The action sequence A^0 is allowed to be coarse and noisy.
[0015] As a further description of the above technical solution:
[0016] The optimization method of the iterative optimization module in step two is as follows: In each round of optimization k:
[0017] a. Add noise: Add controllable noise to some or all of the actions in the current sequence A^{k-1}, or randomly mask some of the actions to obtain a noisy sequence Ã^{k-1};
[0018] b. Parallel decoding and global attention: The non-causal Transformer decoder receives a noisy sequence ∂^{k-1} and an environmental observation O;
[0019] c. Noise reduction prediction: The non-causal Transformer decoder outputs the correction amount of the complete sequence in parallel based on the global path sequence, and obtains a new sequence A^k after one round of optimization.
[0020] d. Condition judgment: Repeat until the preset number of rounds N is reached, or the sequence change is less than a certain threshold.
[0021] As a further description of the above technical solution:
[0022] Step three also includes path replanning when environmental observation O does not match the prediction. The planning method is as follows:
[0023] When the system reaches step t, if the new environmental observation O_{new} does not match the expectation, the system can immediately restart the iterative optimization process in step 2 for the remaining sequence [a_{t+1}^N ... a_T^N], using the currently executed actions [a_1^N ... a_t^N] as constraints.
[0024] As a further description of the above technical solution:
[0025] During the decoding process of the noisy sequence Ã^{k-1}, the non-causal Transformer decoder performs full attention on the noisy sequence Ã^{k-1} in parallel. That is, each action vector is obtained by weighted summation of other action vectors in the sequence, and the weights are the attention scores.
[0026] As a further description of the above technical solution:
[0027] Step a of the iterative optimization module may be skipped and proceed directly to step b, depending on the current coarsness and noise level of A^{k-1}.
[0028] Compared with the prior art, the advantages of this invention are:
[0029] (1) In this scheme, through the global attention mechanism, the model can perceive the global information of the entire sequence when generating each action, thereby generating a highly consistent and smooth action sequence in time, fundamentally avoiding motion jitter caused by local decision-making.
[0030] (2) In this scheme, through multiple iterations of optimization, the system can perform multiple rounds of optimization on the initial rough prediction, gradually correct errors and refine actions, and finally output accuracy is much higher than the result generated by a single autoregression.
[0031] (3) This scheme breaks the drawback of irreversible generation of autoregressive model sequences. Whether in the generation stage or the execution stage, the system can re-optimize the unexecuted part of the sequence, so that it can flexibly cope with environmental changes and uncertainties.
[0032] (4) Although this scheme iterates in multiple rounds, the entire sequence is processed in parallel within each round, avoiding the sequential waiting of autoregression. With sufficient computing resources, the total generation time may be better than the autoregressive generation of long sequences. Attached Figure Description
[0033] Figure 1 This is a schematic diagram of the process of the present invention;
[0034] Figure 2 This is a schematic diagram of the architecture of the iterative optimization module of the present invention. Detailed Implementation
[0035] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.
[0036] Please see Figures 1-2 The present invention provides Embodiment 1:
[0037] A dynamic path correction method for an artificial intelligence robot, comprising:
[0038] Step 1: Initial Sequence Generation. The correction system receives initial environmental observations O and quickly generates a preliminary complete action sequence A^0 = [a_1^0, a_2^0, ..., a_T^0] through an initial predictor. The initial environmental observations O received by the correction system include environmental images and robot states. The action sequence A^0 is allowed to be coarse and noisy.
[0039] The initial sequence can be very coarse. The iterative process gradually converges it to a high-quality, feasible solution. The system does not aim to generate a perfect sequence in one step, but rather improves the sequence quality through multiple rounds of "adding noise and denoising" iterations. The purpose of adding noise is to escape local optima. It allows the model to explore new possibilities different from the current sequence and avoids getting stuck in suboptimal solutions.
[0040] Step 2: Iterative optimization loop. The system does not directly execute the action sequence A^0, but instead sends the action sequence A^0 into an iterative optimization module for N rounds of optimization, where N is greater than 1.
[0041] The optimization method of the iterative optimization module, in each round of optimization k:
[0042] a. Add noise: Add controllable noise to some or all actions of the current sequence A^{k-1}, or randomly mask some actions to obtain a noisy sequence Ã^{k-1}; Step a of the iterative optimization module can be skipped and proceed directly to step b depending on the coarsness and noise level of the current A^{k-1};
[0043] b. Parallel decoding with global attention: The noncausal Transformer decoder receives a noisy sequence ∂^{k-1} and an environmental observation O.
[0044] During the decoding process of the noisy sequence Ã^{k-1}, the non-causal Transformer decoder performs full attention on the noisy sequence Ã^{k-1} in parallel. That is, each action vector is obtained by weighted summation of other action vectors in the sequence, and the weights are the attention scores.
[0045] c. Noise reduction prediction: The non-causal Transformer decoder is based on the global path sequence and outputs the correction amount of the complete sequence in parallel to obtain a new sequence A^k after one round of optimization.
[0046] d. Condition judgment: Repeat until the preset number of rounds N is reached, or the sequence change is less than a certain threshold.
[0047] The full attention added by the non-causal Transformer decoder in step b is achieved by internally transforming an action sequence into a set of vectors and calculating attention scores: for each action in the sequence, the model calculates its attention score with all actions in the sequence.
[0048] For example, the correlation between a2 and a1 may be high because a1 is the previous step; the correlation between a2 and a5 may also be high because a5 is a key obstacle avoidance action, and a2 needs to prepare for it in advance, such as changing direction in advance; the correlation between a2 and a10 may be low because the two are too far apart in time.
[0049] Generate new vectors containing global information: The non-causal Transformer decoder performs a weighted summation of the vectors for all actions based on these attention scores, for example:
[0050] The new vector of a2 = (the vector of fraction 1 × a1) + (the vector of fraction 2 × a2) + (the vector of fraction 3 × a3) + ... + (the vector of fraction T × aT);
[0051] The newly generated vector is no longer just information of a2 itself, but a vector that incorporates the contextual information of the entire sequence. The model can resolve the conflict between short-term interests and long-term goals. The action at the beginning of the sequence can be directly affected by the goal at the end of the sequence because each action knows what the actions before and after it should do. Therefore, the model can generate a very coherent sequence of actions without abrupt changes, avoiding robot jitter. Most importantly, this process is not sequential, but parallel. The model performs the above operation for every action in the sequence at the same time, and all actions gain a global view.
[0052] Step 3: Sequence execution. After N rounds of optimization, the final optimized sequence A^N is output, and the robot begins to execute the actions in the A^N sequence.
[0053] Please see Figure 1 and Figure 2 Based on Example 1, the present invention also provides Example 2:
[0054] If the environmental observation O does not match the prediction during the sequence execution, path replanning is performed. The planning method is as follows:
[0055] When the system reaches step t, if the new environmental observation O does not match the expectation, it can immediately restart the iterative optimization process in step 2 on the remaining sequence [a_{t+1}^N ... a_T^N], using the currently executed actions [a_1^N ... a_t^N] as constraints, to achieve dynamic response and online correction.
[0056] By treating already executed actions as fixed constraints, and only iteratively optimizing future action sequences, the results of previous planning are fully utilized. This makes the computational cost of replanning far less than that of global replanning, achieving true dynamic response.
[0057] Because the optimization starting point is a continuation of the original plan rather than random initialization, the iterative optimization process can usually converge to a high-quality corrected solution very quickly (possibly within 1-2 rounds), meeting the real-time requirements of robot control. It does not discard all previous planning results; the executed parts are fixed as constraints, and the unexecuted parts are used as initial values for optimization, ensuring the efficiency and smoothness of replanning. Globally consistent replanning: this is another manifestation of the power of the global attention mechanism. When optimizing a new sequence [a_{t+1}^{new}, ...], the model can simultaneously consider information from all remaining time steps.
[0058] For example, to avoid colliding with a pedestrian who suddenly appears, the model may not only adjust a_{t+1} (immediate braking), but also coordinate the adjustment of a_{t+3} (early slight turning) and a_{t+5} (alternate path), thereby generating a new path that is safer and more energy-efficient overall, rather than just making an emergency, local response that may cause subsequent problems. By incorporating the latest obstacle information into O, the iterative optimization process will naturally treat it as a new constraint and find a sequence that satisfies the new constraint during the denoising process.
[0059] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.
Claims
1.A dynamic path correction method of an artificial intelligence robot, characterized by, The method comprises: Step 1, initial sequence generation, the correction system receives initial environment observation O, and generates a preliminary complete action sequence A^0 = [a_1^0, a_2^0,..., a_T^0] through an initial predictor; Step 2, iterative optimization loop, the system does not directly execute the action sequence A^0, but sends the action sequence A^0 into an iterative optimization module for N rounds of optimization, where N is greater than 1. Step 3, sequence execution, after N rounds of optimization, the final optimized sequence A^N is output, and the robot starts to execute the actions in the sequence A^N. 2.The dynamic path correction method of an artificial intelligence robot according to claim 1, wherein: The initial environment observation O received by the correction system in step 1 includes environment images and robot states. 3.The dynamic path correction method of an artificial intelligence robot according to claim 1, characterized in that: The action sequence A^0 can be rough and noisy. 4.The dynamic path correction method of an artificial intelligence robot according to claim 1, wherein: In step 2, the optimization method of the iterative optimization module is as follows, in each round of optimization k: a. Add noise, add controllable noise to part or all of the actions of the current sequence A^{k-1}, or randomly mask part of the actions, to obtain a noisy sequence Ã^{k-1}; b. Parallel decoding and global attention, the non-causal Transformer decoder receives the noisy sequence Ã^{k-1} and the environment observation O; c. Denoising prediction, the non-causal Transformer decoder outputs the correction amount of the complete sequence based on the global path sequence in parallel, to obtain a new sequence A^k after one round of optimization; d. Conditional judgment: loop until the preset number of rounds N is reached, or the sequence change is less than a certain threshold. 5.The dynamic path correction method of an artificial intelligence robot according to claim 1, wherein: In step 3, it also includes path replanning when the environment observation O does not match the prediction, and the planning method is as follows: When the tth step is executed, the new environment observation O_{new} does not match the expectation, and the system can immediately restart the iterative optimization process in step 2 for the remaining sequence [a_{t+1}^N... a_T^N] with the constraint of the current executed actions [a_1^N... a_t^N]. 6.The dynamic path correction method of an artificial intelligence robot according to claim 4, characterized in that: In the decoding process of the non-causal Transformer decoder on the noisy sequence Ã^{k-1}, complete attention is performed on the noisy sequence Ã^{k-1} in parallel, that is, each action vector is obtained by weighted summation of other action vectors in the sequence, and the weight is the attention score. 7.The dynamic path correction method of an artificial intelligence robot according to claim 4, characterized in that: According to the roughness and noise level of the current A^{k-1}, step a of the iterative optimization module allows not to directly enter step b.