A curated collection of tested open-source projects for humanoid and legged robots. This repository focuses on practical deployment and engineering experience, covering simulation, reinforcement learning, motion imitation, sim-to-real, and control optimization.
- RL Locomotion Frameworks
- Motion Imitation (AMP/DeepMimic)
- Robotics Toolkits & Optimization
- Evaluation Criteria
Most frameworks in this category are derived from or inspired by the Legged Gym and Isaac Gym ecosystems.
| Project | Description | Rating | Link |
|---|---|---|---|
| Unitree RL Gym | Standard RL for Unitree robots. | ⭐⭐⭐ | GitHub |
| Humanoid-Gym | Optimized for humanoid structures. | ⭐⭐⭐ | GitHub |
| EngineAI Legged Gym | Modified for specific engineai hardware. | ⭐⭐⭐ | GitHub |
- Holosoma ⭐⭐⭐⭐⭐
- Status: Successfully deployed on custom hardware.
- Pros: Excellent sim-to-real performance; robot can walk stably.
- MujocoPlayground ⭐⭐
- Status: Tested.
- Cons: Not friendly for custom robot, bad trainning performance.
- Legged_Lab Extension ⭐⭐⭐⭐
- Features: Supports Sim2Sim and Sim2Real (highly extendable).
- Verdict: Good balance between simulation and real-world potential.
- MimicKit ⭐⭐⭐
- Verdict: Simulation only; lacks direct deployment pipelines.
- BeyondMimic (under test)
- Features: Whole-body tracking in simulation.
- EngineAi AMP (under test)
- Features: Adversarial Motion Priors implementation.
- AMP_for_hardware (under test)
- Features: Classical Adversarial Motion Priors implementation.
- Mujocolab (under test)
- Features: Combines Isaac Lab's manager-based API with MuJoCo Warp.
- GMR (General Motion Retargeting) ⭐⭐⭐⭐⭐
- Features: Supports any humanoid; successfully tested on custom robot platforms.
- Verdict: Highly recommended for motion retargeting.
- Pinocchio ⭐⭐⭐⭐
- Purpose: High-performance Rigid Body Dynamics (FK, IK, etc.).
- Verdict: Industry standard for analytical control.
- Pace-Sim2real ⭐⭐⭐⭐
- Purpose: Actuator parameter identification.
- Verdict: Essential for closing the gap between sim and real actuators.
The ratings are purely based on our hands-on experience with a custom humanoid platform. We focused on:
- Ease of Integration: How difficult is it to swap a custom URDF?
- Sim-to-Real Transfer: Do the policies work on physical hardware without heavy tuning?
- Extensibility: Can we easily modify the reward functions or observation space?
- Motion Quality: Is the gait natural or "jittery"?
“Engineering is about making things work in the real world, not just in a paper.”