AgiBot X1 is a modular humanoid robot designed for research and development, built upon the AimRT framework and utilizing reinforcement learning for control. This repository provides the code specifically for training the reinforcement learning models used by AgiBot X1. The code allows training in simulation, which can then be transferred to the physical robot, or adapted for use with other robot models. The primary approach uses a model-free reinforcement learning algorithm.
This project offers a streamlined workflow for training and evaluating reinforcement learning policies for humanoid robots. It incorporates a modular design allowing for easy integration of new robot models and environments. The code includes utilities for generating JIT and ONNX models for deployment. Detailed documentation and examples facilitate rapid experimentation and adaptation.
- Locomotion Control: Enables training for various legged locomotion tasks like standing and walking.
- Simulation Support: Compatible with Isaac Gym for simulated training and testing.
- Model Export: Provides tools to export trained models to JIT and ONNX formats for deployment.
- Extensible Environments: Supports creating custom robot environments and configurations.
- Scripted Training: Offers flexible training scripts with configurable parameters.
- Sim2Real Validation: Includes tools for validating policies in a simulated environment using Mujoco.
- Joystick Control: Provides a joystick interface for controlling the robot during validation and debugging.
The project has been actively developed and is currently operational, with recent commits and ongoing issue resolution. Comprehensive documentation is provided, but community support is still developing. The code is relatively stable and has been tested with various environments and configurations. However, continuous development is ongoing.
This project serves researchers and developers working with humanoid robotics and reinforcement learning. It's valuable for those seeking to implement intelligent locomotion control for robots, benefiting from a well-structured codebase, extensive documentation, and simulation capabilities. It provides a clear path from training to deployment on real-world robots or integrating with other platforms.