China’s tennis humanoid demo
A China‑developed fully autonomous humanoid tennis robot demonstrated agile footwork, precise swings and sustained rallies, showcasing real‑time perception and decision‑making in dynamic, fast interactions. The demo underscores progress on perception‑to‑action pipelines for agile bipedal systems. (x.com)
The system, called LATENT, is a collaboration led by researchers at Tsinghua University with engineering support from Galbot (Galaxy General Robotics); the project paper lists Zhikai Zhang and colleagues as lead authors. (arxiv.org) Training used a deliberately small dataset: roughly five hours of motion-capture fragments from five amateur players recorded inside a compact 3×5‑meter capture area, not full-match recordings. (arxiv.org) A simulation-first pipeline (MuJoCo) produced policies that were transferred to a real Unitree G1 humanoid for on-court tests, confirming sim‑to‑real deployment on that off‑the‑shelf platform. (github.com/GalaxyGeneralRobotics/LATENT) (newatlas.com) Reported real-world numbers include a 90.9% forehand return rate and roughly 77–78% backhand success in on-court trials, sustained rallies exceeding 25 shots against a human, and real‑court ball speeds reported above 15 m/s. (yahoo.com) Simulation experiments in the authors’ results reached as high as ~96% success on specific shot classes, and the team released code and a tracking sub‑dataset on GitHub to support replication and follow‑on work (tracking code announced March 13, 2026). (newsbreak.com) (github.com/GalaxyGeneralRobotics/LATENT) The deployed tests used external high‑precision motion capture and dedicated ball‑tracking for low‑latency state estimation rather than relying on onboard vision sensors, per the paper and project notes. (arxiv.org)
