HoMMI: whole‑body mobile manipulation
New research called HoMMI demonstrates whole‑body mobile manipulation learned from human demonstrations alone, using egocentric vision and a unified controller to hit ~90% success on tasks like laundry and delivery. It’s a concrete step toward scalable, demo‑based training that bridges the embodiment gap without teleoperation. (x.com)
The paper is authored by Xiaomeng Xu, Jisang Park, Han Zhang, Eric Cousineau, Aditya Bhat, Jose Barreiros, Dian Wang, and Shuran Song, and was submitted to arXiv on March 3, 2026. (hommi-robot.github.io) Reported evaluation numbers show 90% success on a laundry task, 85% on a delivery task, and 80% on a tablescape task, with those results outperforming wrist-only, RGB-only, head-only, and no-neck baselines. (arxivlens.com) Modeling details include a diffusion-based hand‑eye policy conditioned on an embodiment‑agnostic visual encoder together with a relaxed head‑action representation used during training. (arxivlens.com) Execution on real robots relies on a constraint‑aware whole‑body controller that realizes hand‑eye trajectories via differential inverse kinematics while respecting robot‑specific kinematic and safety limits. (arxivlens.com) Data collection extends the Universal Manipulation Interface (UMI) with a head‑mounted egocentric RGB and depth sensing rig to capture global context needed for navigation, bimanual coordination, and active perception. (arxivlens.com) The authors demonstrate cross‑embodiment transfer by deploying policies trained on human demonstrations to a different robot with distinct kinematics and appearance and reporting successful task transfer. (arxivlens.com) Project materials are hosted on an official project site with a results video, and a GitHub repo shows the code release is currently under internal review pending public publication. (hommi-robot.github.io)
