Physical intuition modeling method and system for embodied agent based on five-layer consciousness architecture

Abstract

The application discloses a kind of physical intuition modeling method and system of embodied agent based on five-layer consciousness framework.The method divides the cognitive system of agent into five levels of bottom consciousness layer, P space layer, subconscious layer, J space layer and conscious layer;Among them, P space layer is located between bottom consciousness and subconscious, encodes sensor input as physical intuition hidden space representation, and is divided into four orthogonal subspaces to process body schema, physical intuition, causal risk and situational awareness respectively.P space layer uses incremental learning mechanism to continuously optimize, and has three-dimensional space perception ability of Z-axis, X-axis and Y-axis.When any dimension risk exceeds threshold, it triggers detour decision.The application solves the technical problems that existing embodied agent lacks structured physical intuition module and real-time risk perception capability.

Description

Technical Field

[0001] This invention belongs to the field of embodied AI and cognitive computing technology, specifically relating to a physical intuition modeling method and system that simulates the consciousness structure of a biological organism, used to enhance the interaction, perception and decision-making capabilities of embodied intelligent agents in the physical environment. Background Technology

[0002] Currently, embodied intelligent systems rely primarily on the following three technological approaches for environmental perception and decision-making:

[0003] The first category is decision-making systems based on large language models. These systems convert sensor data into textual descriptions, input them into a large language model for semantic understanding and reasoning, and output action commands. However, large language models are essentially statistical reasoning systems based on textual symbols, lacking the ability to model the causal relationship of the physical world.

[0004] The second type is end-to-end reinforcement learning systems. These systems directly map sensor inputs into control signals through extensive trial and error in environmental interactions. However, their drawbacks include extremely high training costs, requiring millions of environmental interactions; weak generalization ability, necessitating retraining after changes in environmental parameters; and an internal representation that is a "black box," making it impossible to explain the decision-making process.

[0005] The third type is based on world models. These systems model the environment in a latent space and plan behavior through simulations within that latent space. However, their drawback is that although an environmental model is constructed, "physical intuition" is not structurally defined and implemented as an independent functional module. The latent space is singular and homogeneous, and is not decomposed into subspaces according to physical laws.

[0006] In summary, the common problem with existing technologies is the lack of a structured physical intuition processing module, which makes it difficult for intelligent agents to develop an intuitive understanding of physical laws similar to that of biological organisms. Summary of the Invention

[0007] The technical problem to be solved by this invention is: how to build a structured physical intuition module for an embodied intelligent agent, so that it can gradually "grow" an intuitive understanding of physical laws through continuous interaction with the physical environment, and have the ability to perceive risks and make decisions in three-dimensional space.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] This invention divides the cognitive system of an intelligent agent into five levels, from bottom to top: the subconscious layer, the P-space layer, the subconscious layer, the J-space layer, and the conscious layer. Among them, the P-space layer is the core innovative module of this invention, located between the subconscious and the subconscious.

[0010] The mathematical definition of the P-space layer is the mapping function F. P : S ( t - T : t → Z ( t ),in Z ⊆R dP The P-space layer is divided into four orthogonal subspaces: the body schema subspace. Z body Physical Intuition Subspace Z physics Causal risk assessment subspace Z causal and context-aware subspace Z context .

[0011] The P-space layer is updated using an incremental learning approach, with the update rule being: Z ( t +Δ t )= Z ( t )+ η ⋅∇ L (F P ( S ( t - T : t )), S feedback ( t +Δ t )).

[0012] The P-space layer possesses three-dimensional spatial perception capabilities, simultaneously sensing the Z-axis (the firmness of the foot support surface), the X-axis (whether the path ahead is blocked), and the Y-axis (whether there is sufficient space for passage on both sides). When the risk coefficient in any dimension exceeds the threshold, a detour decision is triggered.

[0013] This invention also provides a neural network implementation architecture for the P-space layer, which adopts a three-segment structure: multimodal temporal encoder → four-subspace decomposition expert head → dynamic predictor.

[0014] Compared with existing technologies, the present invention has the following advantages: First, the physical interaction capability of the intelligent agent is improved through the P-space modular design; second, the physical intuition module has the "growth" characteristic through the incremental learning mechanism, and can self-optimize in continuous interaction with the environment; third, the modular design is easy to transfer to different types of intelligent agent platforms; fourth, the three-dimensional spatial perception mechanism has universality and can uniformly handle various spatial obstacles; fifth, the effectiveness and feasibility of the above technical solutions are verified through simulation experiments. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the five-layer consciousness architecture model of the present invention;

[0016] Figure 2 This is a diagram of the P-space neural network implementation architecture of the present invention;

[0017] Figure 3 This is a graph showing the decrease in P-space prediction error in Embodiment 1 of the present invention;

[0018] Figure 4 This is a comparison chart of the navigation performance of the direct-connect architecture and the complete five-layer architecture in Embodiment 2 of the present invention;

[0019] Figure 5 This is a comparison diagram of the startle triggering with and without 3D perception in Embodiment 3 of the present invention;

[0020] Figure 6 This is a diagram showing the verification results of the universality of three-dimensional spatial perception in Embodiment 4 of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] Example 1: Verification of the growth of a single module in P-space

[0023] This embodiment is conducted in a two-dimensional bouncing ball simulation environment. The agent continuously interacts with the physical environment. The input of the P-space module is a sequence of historical motion states, and the output is a prediction of the state at the next moment.

[0024] Simulation results show that the prediction error of the P-space module decreased continuously from an initial value of approximately 32 to below 6.5. This process demonstrates that the P-space module can "grow" physical prediction capabilities through continuous interaction with the physical environment. Figure 3 The curve showing the decrease in prediction error is presented.

[0025] Example 2: Verification of the effectiveness of a five-layer architecture concatenation

[0026] This embodiment is conducted in a two-dimensional navigation task. A complete five-layer consciousness architecture and a direct-connection architecture stripped of the P-space and subconscious layers are configured respectively, and the number of steps required to complete the navigation task is compared between the two.

[0027] Experimental results show that the direct-connect architecture requires 72 steps to complete navigation, while the complete five-layer architecture requires only 63 steps. This result proves that the P-space layer and the subconscious layer play a crucial mediating role in the causal chain, effectively shortening the decision-making path. Figure 4 The trajectory comparison of the two architectures is shown.

[0028] Example 3: Surprise Response and Threshold Mechanism in 3D Perception

[0029] This embodiment is performed in a two-dimensional walking task. The agent needs to traverse a path that includes trap areas.

[0030] When the 3D perception capability in P-space is not enabled, the system cannot predict traps, triggering 24 startle reactions during the walk, and ultimately failing to reach the target. When the 3D perception capability is enabled, the system can identify risks along the Z-axis (foot support), make judgments based on thresholds, and make reasonable detour decisions, successfully passing through the trap area and reaching the target after only 2 detours. Figure 5 It shows a comparison of the number of times surprise is triggered.

[0031] Example 4: Verification of the universality of three-dimensional spatial perception

[0032] This embodiment is carried out on a complex route containing three different spatial obstacles: traps (Z-axis risk), wall obstructions (X-axis risk), and narrow paths (Y-axis risk).

[0033] The P-space layer simultaneously perceives three-dimensional spatial information, and the agent ultimately reaches the target successfully in 99 steps and 10 detours. This result verifies that the P-space three-dimensional perception mechanism can uniformly handle different types of spatial obstacles and has good versatility. Figure 6 It demonstrates the decision-making process of three-dimensional spatial perception.

Claims

1. A physical intuition modeling method for embodied intelligent agents based on a five-layer consciousness architecture, characterized in that, Includes the following steps: Multimodal data of the environment is acquired through sensors and input into the lower consciousness layer; The P-space layer encodes the multimodal sensor data output from the lower consciousness layer into a physical intuition latent space representation. Z ( t The P-space layer is updated using an incremental learning method, and continuous optimization of physical intuition is achieved through prediction error feedback. Furthermore, the P-space layer possesses three-dimensional spatial perception capabilities, enabling it to simultaneously perceive the foot support state corresponding to the Z-axis, the forward path state corresponding to the X-axis, and the spatial states on both sides corresponding to the Y-axis. When the P-space layer detects that the risk coefficient in any dimension exceeds a preset threshold, a detour decision is triggered; The subconscious layer's physical intuition latent space representation Z(t) output from the P-space layer. Z ( t Conduct a survival value assessment; The J-space layer performs causal reasoning and action planning based on the evaluation results of the subconscious layer. The conscious layer transforms the reasoning results of the J-space layer into action instructions and outputs them to the executor.

2. The method according to claim 1, characterized in that, The P-space layer contains four orthogonal subspaces, namely: body schema subspace Z body Physical Intuition Subspace Z physics Causal risk assessment subspace Z causal and context-aware subspace Z context .

3. The method according to claim 1, characterized in that, The encoding mapping function of the P-space layer is expressed as: F P : S ( t - T : t → Z ( t ) in, S ( t )={ s 1( t ), s 2( t ),…, sn ( t )} represents the multimodal sensor input stream of the lower consciousness layer. T For time window, Z ( t ) represents the state prediction vector at the current moment.

4. The method according to claim 1, characterized in that, The update rule for incremental learning is expressed as follows: Z ( t +Δ t )= Z ( t )+ η ⋅∇ L (F P ( S ( t - T : t )), S feedback ( t +Δ t )) Among them, Sfeedback Sfeedback L is the feedback signal after actual physical interaction. L Let η be the prediction error loss function. η This is the learning rate.

5. The method according to claim 1, characterized in that, In the aforementioned three-dimensional spatial perception mechanism, the Z-axis senses the firmness of the foot support surface, the X-axis senses whether the path ahead is blocked, and the Y-axis senses whether there is enough space for passage on both sides.

6. The method according to claim 1, characterized in that, The P-space layer is implemented using a three-segment neural network architecture, which consists of a multimodal temporal encoder, a four-subspace decomposition expert head, and a dynamic predictor.

7. A physical intuition modeling system for embodied intelligent agents based on a five-layer consciousness architecture, characterized in that, Include: Sensor modules are used to acquire multimodal environmental data; A bottom-sensing module is used to receive data from the sensor module; The P-space module is used to encode the data from the bottom consciousness module into a physical intuition latent space representation, and has the ability to perceive three-dimensional space and incremental learning. The subconscious module is used to assess the survival value of the output of the P-space module. The J-space module is used for causal reasoning and action planning based on the assessment results of the subconscious module. The conscious module is used to convert the reasoning results of the J-space module into action commands; An actuator module is used to execute the action instructions.

Images