Galbot hits 96.5% in tennis
China’s Galbot G1 used the LATENT AI framework to reach 96.5% accuracy in tennis rallies after roughly five hours of data, underscoring fast sim‑to‑real motor learning in dynamic tasks. The result raises fresh questions about whether humanoids can rapidly learn high‑speed, contact‑rich skills once thought uniquely human. (interestingengineering.com)
The LATENT paper, titled "Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data," lists Zhikai Zhang and coauthors from Tsinghua University, Peking University, Galbot Inc., Shanghai Qi Zhi Institute and Shanghai AI Laboratory and was posted to arXiv on March 13, 2026. (arxiv.org) The team deployed their learned policy on a Unitree G1 humanoid and reported real‑court strikes against incoming balls with peak velocities above 15 m/s. (arxiv.org) Training relied on short, fragmentary motion‑capture clips collected in a compact 3×5 m capture area and a pipeline that pretrains motion trackers, performs online distillation, and trains high‑level policies in MuJoCo with multi‑GPU simulation. (interestingengineering.com) The authors’ real‑world evaluation averaged results over thousands of on‑court trials (news coverage cites an evaluation set on the order of 10,000 runs) and demonstrated stable multi‑shot rallies against human partners. (letsdatascience.com) Post‑demo coverage reports breakdowns from on‑court tests showing approximately 90.9% success on forehand returns, roughly 77% on backhands, and demonstrations of sustained rallies exceeding 25 consecutive shots. (gadgetreview.com) The team released code, a project page, and a tracking subset on GitHub, with the repository noting a March 13, 2026 update that published the tracking codebase and a small subset of human tennis motion data. (github.com)
