Robot Gait Transition Control With Policy Switching Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for legged robots require retraining of both low-level and high-level controllers for adding new locomotion gaits, leading to limited scalability, increased time, and computation costs, making it difficult to expand the gait library without affecting existing gaits.
Innovation Solution
A transition method that uses a meta-controller and transition-net to enable scalable and iterative deployment of locomotion gaits, allowing for independent policy training and instantaneous switching between gait pairs, leveraging domain randomization and reinforcement learning to ensure robustness and stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple locomotion gaits are employed by way of combination with hierarchical controllers, then the robot can handle challenging terrain and dynamic environments, but the scalability is limited and retraining is required when adding new gaits
Solution Approach 1:
The controller is segmented into independent gait policies, each trained separately to perform specific locomotion tasks. These modular policies can be independently developed and combined without requiring retraining of the entire control system, thus improving scalability while maintaining versatility.
Solution Approach 2:
A universal policy selection mechanism is introduced that can dynamically choose among multiple pre-trained gait policies based on the current task requirements. This universal selector enables the system to handle diverse locomotion scenarios without requiring dedicated controllers for each gait type.
2Reliability
If hierarchical controllers with tight coupling are used to specialize in different gaits, then the robot can perform specific locomotion tasks, but the time cost and computation power cost increase when adding new gaits
Solution Approach 1:
Multiple gait policies are pre-trained independently on various locomotion tasks before deployment. This preliminary action allows the system to have a ready library of specialized gaits that can be quickly selected and executed without requiring time-consuming retraining when new tasks arise.
Solution Approach 2:
Instead of retraining the entire hierarchical controller when adding new gaits, the system copies and utilizes pre-trained policy modules. Each gait policy is a self-contained neural network that can be independently copied and integrated into the system, significantly reducing the time and computational resources required for expansion.
3Reliability
If hierarchical controllers with tight coupling are used to specialize in different gaits, then the robot can perform specific locomotion tasks, but the computation power cost increases when adding new gaits
Solution Approach 1:
The control architecture is segmented into independent, lightweight policy modules that can be selectively activated. Each policy is a compact neural network trained for a specific gait, reducing the overall computational burden compared to a single large hierarchical controller that must handle all gaits.
Solution Approach 2:
The system dynamically changes operational parameters by selecting different pre-trained policies based on task requirements. This parameter switching approach avoids the need to maintain and compute with a single complex controller that must adapt to all possible gaits, thereby reducing computation power consumption.
Data Source
AI summary
A transition method of locomotion gait of a robot includes: executing a deployment procedure multiple times, each execution includes: randomly selecting a source policy and a destination policy, simulating a transition operation from the source policy to the destination policy, and recording a transition configuration and a transition result to a transition database, where each policy is a neural network model, and a latent state in the transition configuration is a hidden layer of the neural network model of the source policy. The method further includes: training a transition-net according to the transition database, and performing the following steps by a meta-controller disposed on the robot: selecting two gait policies as an active policy and a queued policy, executing the active policy, inputting the two policies to the transition-net to obtain a success probability, and when the success probability is greater than a threshold, executing the queued policy.