Proposed Enhancement: Reinforcement Learning Module Integration
This proposal aims to add a reinforcement learning (RL) module that provides interfaces for directly deploying RL algorithms within the framework. The goal is to enable users to easily design custom reward functions and train LeRobot manipulation policies in the LeIsaac environment.
In addition, I hope to integrate this module with the existing state-machine-based automatic data collection pipeline.
The initial plan is to use the rsl_rl library to implement the PPO algorithm. In future iterations, support for additional RL algorithms such as SAC and TD3 is expected.
I am still exploring the design. At the moment, it seems that an imitation learning (IL) module may be easier to implement and more naturally integrated with the data generation pipeline. Once the plan becomes more mature, I will provide a more detailed proposal.
Proposed Enhancement: Reinforcement Learning Module Integration
This proposal aims to add a reinforcement learning (RL) module that provides interfaces for directly deploying RL algorithms within the framework. The goal is to enable users to easily design custom reward functions and train LeRobot manipulation policies in the LeIsaac environment.
In addition, I hope to integrate this module with the existing state-machine-based automatic data collection pipeline.
The initial plan is to use the
rsl_rllibrary to implement the PPO algorithm. In future iterations, support for additional RL algorithms such as SAC and TD3 is expected.I am still exploring the design. At the moment, it seems that an imitation learning (IL) module may be easier to implement and more naturally integrated with the data generation pipeline. Once the plan becomes more mature, I will provide a more detailed proposal.