Humanoid learns tennis fast
Researchers trained a humanoid to play multi-shot tennis using roughly five hours of motion-capture data and achieved ~90% success on balls over 15 m/s in rallies with humans reported. The work — described in more technical coverage as LATENT — shows robots can learn high-speed manipulation from imperfect human data, accelerating sim-to-real skill transfer reported.
The research was posted as a preprint on arXiv) on March 13, 2026 and lists contributors from Tsinghua University, Peking University, Galbot Inc., Shanghai Qi Zhi Institute, and Shanghai AI Laboratory. (arxiv.org) LATENT’s control stack centers on a learned latent action space plus a hierarchical high-level policy and a dedicated wrist-correction module to compose fragmented human motions into full strokes. (arxiv.org) The authors simulated training in MuJoCo and report a deployment pipeline designed for robust sim‑to‑real transfer before running the learned policy on an actual humanoid. (github.com) Real‑world trials used the Unitree G1 humanoid platform for physical testing, with external motion‑capture systems providing low‑latency ball and human tracking rather than onboard cameras. (arxiv.org) The team published an official implementation and noted a March 13, 2026 release of tracking code and a small subset of their tennis motion data on GitHub, making parts of the pipeline reproducible. (github.com) Multiple outlets filmed and reported that LATENT sustained multi‑shot rallies against human players and also demonstrated robot–robot rallying in demos, coverage that highlighted the project’s sim‑to‑real robustness. (robohorizon.com)
