A research team from Tsinghua University, Peking University, and Galbot (GalaxyGeneralRobotics) has released LATENT — a framework that teaches humanoid robots athletic tennis skills entirely from noisy, imperfect human motion data.

The Approach

The system works in three stages: a motion tracker pre-trains on human tennis recordings captured via motion capture, an online distillation phase transfers those skills into a policy for the Unitree G1 robot, and a high-level tennis-playing policy governs real-world shot decisions. Simulation training runs in MuJoCo with multi-GPU support.

The name is an acronym: Learning Athletic humanoid TEnnis skills from imperfect human motioN daTa.

Why This Matters

Most robot athletic training relies on near-perfect motion capture data — expensive, lab-constrained, and hard to scale. LATENT's pipeline tolerates noisy inputs, making it significantly more practical. The Unitree G1 achieves multi-shot rallies with dynamic full-body coordination and rapid reactions, not just static swings.

The demo footage shows the G1 returning shots with agile, fluid movement that mimics real tennis form — a meaningful step toward general-purpose humanoid athletic skills.

Open Source

The team released the tracking codebase and a subset of human tennis motion data on March 13, 2026, with more checkpoints and training components rolling out on GitHub. The framework is designed to generalize beyond tennis to other athletic motions.

Researchers are from Tsinghua University, Peking University, Galbot, Shanghai Qi Zhi Institute, and Shanghai AI Laboratory.