Unitree G1 beats creator at tennis

If imperfect motion fragments work for tennis, this pipeline applies to any dexterous task where clean motion capture is impractical — and that's most real-world robotics.

• –LATENT learns a latent action space from messy, partial human motion data, then uses a high-level policy to correct and compose primitive skills, bypassing the data quality bottleneck that's plagued robotics for years
• –Only ~5 hours of training data required, orders of magnitude less than typical MoCap-based approaches — a signal that data efficiency is now the frontier, not just compute
• –Robust sim-to-real transfer is built in, addressing one of humanoid deployment's hardest unsolved problems
• –Deployed on the Unitree G1, a commercially available platform — this isn't trapped in a lab, it's something startups can pick up today
• –Code and a subset of motion data are publicly released on GitHub, enabling follow-on research without expensive proprietary data collection