New motion‑primitive methods & bimanual dexterity
Researchers surfaced LATENT — a method that combines motion primitives with RL to control humanoids without full demonstration datasets reported — and Stanford published HandelBot advances in sim‑to‑real bimanual manipulation for piano playing reported. Both moves push the field away from purely imitation pipelines toward modular skill primitives and improved sim‑to‑real transfer. The techniques matter for teams building robust, generalizable manipulation on real robots.
LATENT is authored by researchers from Tsinghua University, Peking University, Shanghai Qi Zhi Institute / Shanghai AI Laboratory, and Galbot Inc. (arxiv.org) The LATENT team deployed their policy on a Unitree G1 humanoid robot and demonstrated stable multi‑shot tennis rallies against human players with peak incoming ball speeds above 15 m/s. (arxiv.org) The LATENT codebase documents a pipeline that includes motion‑tracker pre‑training, online distillation into a latent action space, and high‑level policy learning using MuJoCo with multi‑GPU training, with a tracking code release and a small subset of human tennis motion data announced on March 13, 2026. (github.com) HandelBot is credited to Amber Xie, Haozhi Qi, and Dorsa Sadigh (Stanford) and reports that a simulation policy plus structured trajectory refinement and residual RL achieves successful real‑world piano execution after as little as 30 minutes of physical interaction. (arxiv.org) The HandelBot paper reports a 1.8× improvement over direct sim‑to‑real deployment and validates the system across five songs in hardware experiments. (arxiv.org) HandelBot’s implementation repository includes a dg5f_driver and assets for Tesollo DG‑5F five‑finger hands alongside Franka robot descriptions, indicating the real‑world setup uses DG‑5F hands mounted on Franka arms. (github.com) The HandelBot repo also specifies their ManiSkill piano training environment and exact horizon lengths used to reproduce results (e.g., Twinkle Twinkle: 160 timesteps; Ode to Joy: 330; Fur Elise: 320), and both projects publish code and project pages for replication and videos. (github.com)
