Humanoid Robot Tennis: LATENT Framework Teaches Athletic Skills from Imperfect Motion-Capture Data

Researchers have published LATENT, a framework for training full-size humanoid robots to perform athletic tennis skills by learning from imperfect, incomplete human motion-capture data rather than requiring expert demonstrations. The system addresses a core bottleneck in humanoid locomotion research — the scarcity of high-quality reference data for complex, whole-body athletic motions — by enabling policy learning that tolerates noise and missing keypoints. Evaluated on a humanoid platform, LATENT produces physically plausible and transferable tennis swing and serve policies directly from noisy mocap sequences.

Key Takeaways

• LATENT tolerates imperfect mocap data — missing markers, noise, and incomplete sequences — removing expert-demo bottleneck for athletic humanoid training
• Framework demonstrated on full humanoid with tennis-specific whole-body coordination (swing, serve, lateral footwork)
• Paper published March 15, 2026; project page at zzk273.github.io/LATENT with videos and supplementary results

Original source: Hacker News / LATENT Project