hanmingbai / mimic_distill_walk

What is mimic_distill_walk?

This Python project uses mimic distillation to achieve natural human-like walk instead of AMP's mode collapse and reward-heavy gaits. Train a teacher policy via imitation on motion datasets like seed_g1, then distill it into a velocity-following student policy deployable on Unitree robots such as G1, Go2, B2, and H1. Users get sim-to-real locomotion via Isaac Lab training and C++ deployment with joystick control.

Why is it gaining traction?

It delivers fluid, human-like motion tracking from datasets, bypassing brittle reward tuning in alternatives like echo mimic github or groot mimic github tools. Key hook: FSM joystick states for passive, standing, velocity walk, even mimic dances (gangnam style), with ONNX export for low-latency real-robot runs. Devs notice robust sim-to-sim transfers and easy policy swaps.

Who should use this?

Unitree robotics engineers training legged walkers on Isaac Lab, especially those distilling mimic data github assets into velocity policies for G1 bipeds. Ideal for researchers mimicking cxr github motions or visual mimic github clips, seeking hardware-ready natural walk over stiff RL baselines.

Verdict

Grab it if you're on Unitree hardware—deploy code shines for quick tests—but 1.0% credibility score and 19 stars signal early maturity; docs guide training well, yet expect tweaks for production. Solid foundation for custom mimic distillation experiments.

(198 words)