An IsaacLab training stack for marathon-winning humanoids

What it does TienKung-Lab is a reinforcement-learning locomotion stack for the full-sized TienKung humanoid. It uses IsaacLab to train walking and running policies with Adversarial Motion Priors and periodic gait rewards, then pushes them through Sim2Sim validation in MuJoCo and onto the physical robot. The project also retargets human motion-capture data from SMPLX datasets into robot trajectories.

The interesting bit This is not a toy sim. The same TienKung hardware that won the first Humanoid Robot Half Marathon runs policies trained here, and the README soberly warns that real-world RL can trigger violent motions—so check your emergency stop. Ray-casting sensors are already wired in for obstacle avoidance, though full perceptive control remains on the TODO list.

Key highlights

• Trains policies in IsaacLab using AMP-style rewards blended with periodic gait rewards.
• Validates trained policies in MuJoCo via Sim2Sim before real-world deployment.
• Retargets SMPLX motion-capture data (AMASS, OMOMO) into robot motion using GMR.
• Integrates ray-casting sensors for environmental perception and obstacle avoidance.
• Proven on the physical TienKung robot; deployment code lives in a separate repository.

Caveats

• Motion retargeting is currently limited to SMPLX-type datasets only.
• Additional sensors and perceptive control are still marked as unfinished.
• The README carries an explicit safety warning: RL policies may cause unexpected or violent motions on real hardware.

Verdict Worth a look if you are building sim-to-real humanoid locomotion and need an IsaacLab reference that has already crossed the reality gap. Otherwise, it is too tightly bound to the TienKung platform and NVIDIA’s stack to be broadly reusable.