Tsinghua's LATENT learns tennis
Tsinghua’s LATENT project trained a humanoid to play dynamic tennis using human demonstration data, advancing whole‑body coordination and fast reactive control for sport‑scale tasks. The work shows progress in learning dynamic, contact‑rich behaviors for humanoid autonomy reported.
The LATENT manuscript was posted to arXiv as arXiv:2603.12686 on March 13, 2026 by Zhikai Zhang et al., listing Tsinghua University, Peking University, Galbot, Shanghai Qi Zhi Institute and Shanghai AI Laboratory as co‑affiliations. (arxiv.org) The public codebase documents a three‑stage pipeline—motion‑tracker pre‑training, online distillation (DAgger), and high‑level policy training—and runs training in MuJoCo/Brax with tools to convert tracking models to ONNX for evaluation. (github.com) Hardware modifications for the demo included a Unitree G1 humanoid platform noted in the paper and a custom 3D‑printed racket adapter plus reflective markers used with an external optical motion‑capture system during real‑world tests. (zzk273.github.io) The project page and PDF show demo conditions with incoming balls on the order of 15 m/s in figure captions, and press reports quote measured return rates—forehands above 90% and backhands around 78%—in the team’s live rallies. (zzk273.github.io) The GitHub README notes a March 13, 2026 release of the tracking codebase and a small subset of retargeted tennis motion data and lists TODOs to publish pretrained latent models, the DAgger distillation codebase, and sim‑to‑real design files. (github.com) Lead author Zhikai Zhang is quoted saying the robot’s performance climbed from “couldn’t return a single ball” on day one to outperforming its developer by the project’s end, an anecdote reported in press coverage of the release. (robohorizon.com)
