GalaxyGeneralRobotics / LATENT

What is LATENT?

LATENT is a Python pipeline for training simulated humanoid robots, like Unitree G1, to mimic imperfect human tennis motions captured via trackers. Using MuJoCo and Brax with JAX for multi-GPU speed, it runs PPO to learn low-level tracking policies, handling domain randomization for robustness. Users download a motion dataset subset, train via simple CLI commands, export to ONNX, and visualize side-by-side robot vs. reference videos—ideal for robotics imitation from mocap data akin to latent diffusion or latent video diffusion techniques.

Why is it gaining traction?

It stands out by open-sourcing a full motion tracking pipeline from a Tsinghua/Galbot paper, including official GitHub actions for reproducible training and a small human tennis dataset. Developers dig the quick setup with uv sync, wandb logging, and eval tools that convert Brax models to ONNX for real-time playback, bridging sim-to-real gaps better than generic RL libs. The latent action focus echoes latent consistency models or latent flow matching, but tailored for athletic humanoid skills.

Who should use this?

Robotics engineers building humanoid locomotion or manipulation from mocap data, especially those targeting sports-like tasks. RL researchers in imitation learning who need MuJoCo baselines with domain rand for tennis swings or similar dynamic motions. Sim teams at labs like Galbot prototyping latent scope behaviors before hardware transfer.

Verdict

Grab it if you're in humanoid RL—solid foundation for tracking, but 1.0% credibility score and 45 stars signal early maturity; docs are README-focused with TODOs for full data and policies. Watch for releases to boost sim2real tennis skills.

(198 words)