We present Humanoid Touch Dream, a framework for dexterous, contact-rich humanoid manipulation.
Our platform integrates VR-based whole-body teleoperation, an RL-based lower-body controller, dexterous hand retargeting, distributed tactile sensing, and a multimodal policy called Humanoid Transformer with Touch Dreaming (HTD). HTD learns from vision, proprioception, force, and tactile signals, and is trained with a touch dreaming objective that predicts future hand-joint forces and future tactile latents for contact-aware representation learning.
We demonstrate the system on several real-world contact-rich tasks, including tight-tolerance insertion, book organization, towel folding, cat litter scooping, bimanual tea serving, and long-horizon lab cleaning.
Learn more about this research:
humanoid-touch-dream.github.io
