Artificial intelligence robot control system fused with world model architecture

Abstract

The invention discloses an artificial intelligence robot control system fused with a world model architecture, and the system comprises a control main system which comprises a shared multi-mode backbone network module, a strategy head, and a world model head. A visual encoder, an ontology perception encoder, a text encoder and multi-mode fusion are integrated in the shared multi-mode backbone network module, and the strategy head is used for generating a current action instruction. According to the method, a self-supervision signal provided by a world model is used as an additional training constraint, the dependence on large-scale teaching data is reduced, and a strategy head generates actions based on representation rich in physical dynamic information, so that the device has the advantages that the decision is more stable when facing environmental noise or uncertainty; error accumulation in a long-range task is remarkably reduced, and compared with a traditional open-loop strategy model, the method has the advantage that the generalization ability of training out-of-distribution scenes is remarkably improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, robot control, and machine learning, and more specifically, to an artificial intelligence robot control system that integrates a world model architecture. Background Technology

[0002] Data-driven policy learning has become the mainstream paradigm for training robots to perform complex tasks. Among them, deep neural networks based on end-to-end training, especially methods represented by vision-language-action (VLA) models, have achieved remarkable success. However, most current methods suffer from a fundamental limitation.

[0003] Existing robot policy models typically follow an open-loop perception-action mapping paradigm. The model inputs are the current observation (such as camera image o_t, ontological perception information p_t) and high-level instructions (such as text l), and the output is the action to be performed a_t, i.e., the learning function: π(o_t, p_t, l) → a_t.

[0004] The flaw in this paradigm is that the action $a_t$ itself is only the model's output, never digested or understood as part of its input. The model is like a "black box," learning "what should be done in this situation" but not truly understanding "why this action works" or "what the world will become after this action is performed." It lacks explicit modeling of the causal relationship between action and state changes, which leads to:

[0005] Insufficient understanding of physical laws: The model has difficulty understanding physical concepts such as force and collision, and has poor generalization ability when encountering objects or scenes outside the training distribution.

[0006] Accumulation of sequential decision-making errors: In long-term tasks, because the long-term consequences of one's own actions cannot be predicted, small decision-making errors will accumulate and eventually lead to task failure.

[0007] The learning process requires a large amount of real-world interaction data because the model cannot simulate trial and error internally.

[0008] World models offer a solution to this problem. The core of a world model is learning environment dynamics, which has the functional form: f(o_t, a_t) → o_{t+1}. It explicitly learns "what changes will occur in the next state o_{t+1} when action a_t is performed in the current state o_t". However, existing work typically trains or uses the world model and policy model separately, failing to achieve deep synergy between the two at the representation learning level.

[0009] Therefore, there is an urgent need in this field for a unified architecture that can deeply integrate policy learning with world model learning, enabling robot policies to have a deep understanding of physics. Summary of the Invention

[0010] To address the problems existing in the prior art, the purpose of this invention is to provide an artificial intelligence robot control system that integrates a world model architecture.

[0011] To solve the above problems, the present invention adopts the following technical solution.

[0012] An artificial intelligence robot control system integrating a world model architecture includes a main control system, which includes a shared multimodal backbone network module, a strategy head, and a world model head.

[0013] The shared multimodal backbone network module integrates a visual encoder, an ontology-aware encoder, a text encoder, and multimodal fusion.

[0014] The strategy header is used to generate the current action instruction;

[0015] The world model head is used to predict the next state after an action is performed.

[0016] As a further description of the above technical solution: the training data of the main control system is in tuple form as follows:

[0017] Current visual observation o_t, ontology perception p_t, text instruction l, action a_t, next visual observation o_{t+1}, next ontology perception p_{t+1}.

[0018] As a further description of the above technical solution: the training process of the main control system is as follows:

[0019] The shared backbone network processes the input to generate the representation h_t;

[0020] The strategy head outputs the predicted action â_t;

[0021] The world model head outputs the predicted state ô_{t+1};

[0022] Hybrid loss function

[0023] As a further description of the above technical solution: the inference stage of the main control system is as follows:

[0024] Input the current observation (o_t, p_t, l), and mask the action input;

[0025] The backbone network extracts h_t, and the policy header outputs the execution action a_t;

[0026] Optional activation of the world model head allows for simulation of the consequences of actions.

[0027] As a further description of the above technical solution: the main control system enables the backbone network to learn the causal relationship between actions and states through state prediction constraints of the world model head, thereby realizing the internalization and encoding of physical laws.

[0028] As a further description of the above technical solution: the shared multimodal backbone network module supports visual, ontology perception, and text multimodal input, and is compatible with Transformer and DiffusionModel architectures.

[0029] As a further description of the above technical solution: the strategy head generates actions based on physical dynamic representation, thereby improving robustness in the face of environmental uncertainties.

[0030] Compared with the prior art, the advantages of this invention are:

[0031] This scheme forces the shared backbone network to learn explicit causal relationships between actions and state changes by constraining the state prediction of the world model head. This makes policy decisions naturally conform to physical common sense, thereby enabling the device to have a significantly improved generalization ability to scenarios outside the training distribution compared to traditional open-loop policy models.

[0032] This scheme uses the self-supervised signal provided by the world model as an additional training constraint to reduce the dependence on large-scale teaching data. The policy head generates actions based on representations rich in physical dynamic information, thereby enabling the device to make more stable decisions when facing environmental noise or uncertainty, and significantly reduce error accumulation in long-term tasks. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the principle block of the present invention.

[0034] Explanation of the labels in the diagram:

[0035] 1. Control main system; 2. Shared multimodal backbone network module; 3. Strategy head; 4. World model head. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention;

[0037] To address the current problems of control systems still struggling to understand physical concepts such as force and collision, and exhibiting poor generalization ability when encountering objects or scenes outside the training distribution, Implementation Example 1 is proposed:

[0038] Please see Figure 1In this embodiment, an artificial intelligence robot control system integrating a world model architecture includes a main control system 1, which includes a shared multimodal backbone network module 2, a strategy head 3, and a world model head 4.

[0039] The shared multimodal backbone network module 2 integrates a visual encoder, an ontology-aware encoder, a text encoder, and multimodal fusion.

[0040] Strategy Header 3 is used to generate the current action instruction;

[0041] The world model head 4 is used to predict the next state after an action is performed;

[0042] The training data for the main control system 1 is in tuple form as follows:

[0043] Current visual observation o_t, ontology perception p_t, text instruction l, action a_t, next visual observation o_{t+1}, next ontology perception p_{t+1};

[0044] The training process of the main control system 1 is as follows:

[0045] The shared backbone network processes the input to generate the representation h_t;

[0046] Strategy head 3 outputs the predicted action â_t;

[0047] The world model's first four outputs predict the state ô_{t+1};

[0048] Hybrid loss function ;

[0049] The reasoning phase of the main control system 1 is as follows:

[0050] Input the current observation (o_t, p_t, l), and mask the action input;

[0051] The backbone network extracts h_t, and the policy header outputs the execution action a_t;

[0052] Optional activation of world model head 4 to simulate the consequences of actions;

[0053] The main control system uses the state prediction constraints of the world model head to enable the backbone network to learn the causal relationship between actions and states, thereby achieving the internalization and encoding of physical laws.

[0054] The shared multimodal backbone network module supports visual, ontology perception, and text multimodal input, and is compatible with Transformer and DiffusionModel architectures.

[0055] In this embodiment, by constraining the state prediction of the world model head 4, the shared multimodal backbone network module 2 is forced to learn the explicit causal relationship between actions and state changes, so that the decision of the policy head 3 naturally conforms to physical common sense. This enables the device to have a significantly improved generalization ability to scenarios outside the training distribution compared to traditional open-loop policy models.

[0056] Furthermore, to address the problem that current systems, in long-term tasks, cannot predict the long-term consequences of their actions, and that small decision-making errors accumulate over time, ultimately leading to task failure, a second implementation example is proposed:

[0057] Please see Figure 1 Among them, strategy head 3 generates actions based on physical dynamic representation, which improves robustness in the face of environmental uncertainty.

[0058] In this invention, the self-supervised signal provided by the world model head 4 serves as an additional training constraint, reducing the reliance on large-scale teaching data. The strategy head 3 generates actions based on representations rich in physical dynamic information, thereby enabling the device to make more stable decisions when facing environmental noise or uncertainty, and significantly reduce error accumulation in long-term tasks.

[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. An artificial intelligence robot control system integrating a world model architecture, comprising a main control system (1), characterized in that: The main control system (1) includes a shared multimodal backbone network module (2), a strategy head (3), and a world model head (4). The shared multimodal backbone network module (2) integrates a visual encoder, an ontology perception encoder, a text encoder, and multimodal fusion. The strategy header (3) is used to generate the current action instruction; The world model head (4) is used to predict the next state after an action is performed.

2. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The training data for the main control system (1) is in tuple form as follows: Current visual observation o_t, ontology perception p_t, text instruction l, action a_t, next visual observation o_{t+1}, next ontology perception p_{t+1}.

3. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The training process of the main control system (1) is as follows: The shared backbone network processes the input to generate the representation h_t; The strategy head outputs the predicted action â_t; The world model head outputs the predicted state ô_{t+1}; Hybrid loss function .

4. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The reasoning phase of the main control system (1) is as follows: Input the current observation (o_t, p_t, l), and mask the action input; The backbone network extracts h_t, and the policy header outputs the execution action a_t; Optional activation of the world model head allows for simulation of the consequences of actions.

5. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The main control system (1) enables the backbone network to learn the causal relationship between actions and states through the state prediction constraints of the world model head (4), thereby realizing the internalization and encoding of physical laws.

6. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The shared multimodal backbone network module (2) supports visual, ontology perception, and text multimodal input, and is compatible with Transformer and DiffusionModel architectures.

7. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The strategy head (3) generates actions based on physical dynamic representation, thereby improving robustness in the face of environmental uncertainties.

8. The artificial intelligence robot control system based on a world model architecture according to claim 1, characterized in that: The main control system (1) reduces its reliance on large-scale teaching data through the self-monitoring signal of the world model head (4).

Images