Touch Dreaming HRI tests
- Researchers released Touch Dreaming results that predict contact during manipulation tasks like insertion and folding. - The method uses simulated and VR-derived contact priors to inform robot touch expectations. - The approach was framed as a way to reduce costly physical trials by predicting contact before execution. ( )
Robots usually learn touch the hard way: by bumping into things over and over. A new system called Touch Dreaming trains a humanoid robot to predict contact before it moves. (arxiv.org) The work appears in an April 2026 paper, “Learning Versatile Humanoid Manipulation with Touch Dreaming,” by researchers from Carnegie Mellon University, the University of Texas at Arlington, and Bosch Center for AI. The team built a Humanoid Transformer with Touch Dreaming, or HTD, that combines touch with camera views and proprioception, the robot’s sense of its own body position. (arxiv.org) In plain terms, the model is trained not just to choose the next action, but also to forecast what the robot’s hands should feel next. The paper says HTD predicts future hand-joint forces and future tactile latents, compressed internal signals that stand in for touch patterns. (arxiv.org) That matters most in tasks where vision is not enough. A robot inserting a part with 3.5-millimeter clearance, folding a towel, or scooping cat litter can lose track of contact from cameras alone, while touch can signal pressure, slip, and alignment. (arxiv.org) The researchers paired that prediction model with a reinforcement-learning whole-body controller and a virtual-reality teleoperation setup for collecting demonstrations. Their project page says the data system maps human motion onto a humanoid robot so operators can show full-body manipulation skills in the real world. (humanoid-touch-dream.github.io) The team tested the system on five contact-heavy tasks: Insert-T, Book Organization, Towel Folding, Cat Litter Scooping, and Tea Serving. Across those tasks, the project page and paper report a 90.9% relative improvement in average success rate over the stronger baseline. (humanoid-touch-dream.github.io, arxiv.org) One ablation result points to the core claim. The authors report that predicting touch in a latent space, instead of predicting raw tactile readings directly, produced a 30% relative gain in success rate. (arxiv.org) The release lands as humanoid robotics groups push beyond pick-and-place demos toward longer, messier household and workplace tasks. The paper frames contact-aware learning as a way to handle frequent contact changes without relying only on costly physical trial-and-error. (arxiv.org) The project site shows autonomous rollouts at original speed, including towel folding and insertion, and notes that some cat-litter videos were recorded while the robot’s left hand had intermittent communication failures. The code repository says code for the controller, teleoperation stack, and HTD policy learning is planned for release by early May 2026. (humanoid-touch-dream.github.io, github.com) The pitch is straightforward: let the robot “imagine” touch before contact happens, then use that expectation to move with less guesswork. The next test is whether those gains hold up as more labs try the method outside the authors’ own setup. (arxiv.org, github.com)
