8 results about "Motion strategy" patented technology

This application relates to the fields of intelligent robot control and artificial intelligence technology, and discloses a robot dynamic trajectory control method based on reinforcement learning and multi-sensor fusion. This method first utilizes extended Kalman filtering to fuse multimodal sensor data and then predicts the temporal distribution of dynamic obstacles through a long short-term memory network. Simultaneously, it calculates the robot's dynamic maneuverability and singularity risk factors. Based on this, a reinforcement learning network based on an attention mechanism is constructed to generate high-level motion strategies, which are then transformed into low-level execution instructions via a variable impedance controller and a convex optimization torque allocation module, utilizing null space characteristics to avoid singularities. Furthermore, the method fine-tunes model parameters in real time by monitoring prediction errors and the degree of safety intervention online. This invention achieves the organic integration of logical decision-making and physical execution through a hierarchical architecture, significantly improving the robot's operational safety, robustness, and control accuracy in dynamic environments.

Systems, methods, and other embodiments described herein relate to estimating a policy for object motion by training a multi-modal model using diffusion through inferred goals and noise. In one embodiment, a method includes training a multi-modal model to generate a policy using semi-labeled data derived from wild data, and the multi-modal model predicts operator intent and a parameter associated with the policy for an agent in motion. The method also includes expanding outputs from the multi-modal model using noise within a diffusion model, and the noise augmenting the semi-labeled data. The method also includes feeding the outputs including the noise and the wild data to the multi-modal model until satisfying a training parameter associated with the policy.

The application discloses a kind of multi-unmanned ship cooperative hunting methods of fusing graph neural network and attention, comprising: constructing hunting scene and initializing pursuit and escape state, establishing unmanned ship kinematics model and target escape movement strategy;Global state and single-boat observation information are obtained, and hunting features are extracted;Graph neural network is constructed to weight aggregation features, and graph aggregation features are output and input into multi-agent reinforcement learning model;Unmanned ship control action is combined with kinematics model to interact with environment, and target state is updated synchronously to obtain new global state, and reward value is calculated according to hunting situation, and relevant data is stored in experience replay pool.Sample meets standard, and strategy network and value network parameters are updated, and loop iteration is repeated until termination condition is met, and trained strategy network is output to realize cooperative hunting.The application introduces multi-head self-attention value network, adapts to marine dynamic escape scene containing island barrier, and improves global situation awareness, key interaction identification, cooperative hunting stability and decision efficiency.

A cleaning robot (10), and a method and apparatus for same returning to a base (20), and a method and apparatus for same moving out of a base (20). On the basis of the type of a target base and target task information, the cleaning robot (10) can be controlled to move onto the target base or move out of the target base according to a corresponding movement strategy, such that the risk of the cleaning robot (10) becoming stuck on the base (20) and being unable to move out, or the risk of the cleaning robot (10) being unable to move onto the base (20) is reduced.

The application discloses a ship unmanned aerial vehicle game guidance and control method for channel ice condition detection, relates to the technical field of ship / unmanned aerial vehicle motion control research, and comprises the following steps: establishing a ship three-degree-of-freedom and unmanned aerial vehicle six-degree-of-freedom nonlinear model; designing a ship expected task point according to an actual route point, and generating an unmanned aerial vehicle task point according to a detection radius of the unmanned aerial vehicle and a water width; taking the task point as an expected coordinate, constructing a local optimal motion strategy between the ship / unmanned aerial vehicle and a virtual body based on a Stackelberg game, and calculating a global optimal motion strategy according to the local optimal motion strategy; designing a position controller, fusing the ship local optimal strategy and real-time position of the unmanned aerial vehicle, and making attitudes of both sides converge through a total controller composed of an attitude virtual controller and an attitude controller; driving the ship and the unmanned aerial vehicle to execute the optimal strategy through a cooperative action of the position and attitude controllers, and realizing cooperative control; and the application can meet the marine engineering demand of high-precision navigation and resistance to marine environment interference.

In various examples, policy prediction-based motion planner systems and methods for autonomous and semi-autonomous systems and applications are provided. A scenario tree structure may be generated that represents potential behaviors of one or more peripheral agents based on perception data of a scene within which an ego vehicle operates. A joint MPC algorithm may optimize the motion of an ego vehicle within the context of the scenario tree structure to produce a policy tree structure. An MPC policy prediction model may be trained to predict the policy tree structures that a joint MPC algorithm would produce, given a set of environmental perception data. An ego vehicle may comprise a trained MPC policy prediction model that receives perception data, and based on that input predicts a policy tree structure that may be used to define a motion policy for navigating the ego vehicle through the scene.

This invention discloses a multi-unmanned surface vessel (USV) collaborative encirclement method integrating graph neural networks and attention, comprising: constructing an encirclement scenario and initializing the pursuit / escape state; establishing a kinematic model of the USV and a target escape motion strategy; acquiring global state and single-USV observation information, and extracting encirclement features; constructing a graph neural network to weighted aggregate the features, outputting the aggregated graph features and inputting them into a multi-agent reinforcement learning model; combining the USV control actions with the kinematic model interaction environment, synchronously updating the target state to obtain a new global state, calculating the reward value based on the encirclement situation, and storing the relevant data in an experience replay pool. After the sample reaches the target, the parameters of the policy network and value network are updated, iterating cyclically until the termination condition is met, and outputting the trained policy network to achieve collaborative encirclement. This invention introduces a multi-head self-attention value network, adapting to dynamic marine escape scenarios with island and reef obstacles, improving global situational awareness, key interaction identification, and the stability and decision-making efficiency of collaborative encirclement.

The application provides a redundant space robot obstacle avoidance planning method based on a relaxed null space, comprising: constructing an end desired trajectory generator based on artificial demonstration; constructing an end desired trajectory tracking controller; constructing a relaxed null space motion strategy of the redundant space robot according to mapping of end motion of the redundant space robot and joint motion; constructing state variables and action variables of the redundant space robot according to parameters related to the redundant space robot and obstacles; constructing an obstacle avoidance reward function of the redundant space robot according to an end relaxed motion vector; constructing a relaxed null space obstacle avoidance motion training strategy according to the variables and functions constructed above; and obtaining a relaxed null space obstacle avoidance motion planner according to the training obstacle avoidance motion planning strategy network and the end desired trajectory generator. According to the technical scheme provided by the embodiment of the application, the redundant space robot can track the end desired trajectory as much as possible while avoiding collision.