Humanoid control paper: HTD

Humanoid robots still struggle with the ordinary act of touching things while they move, and a new paper from Carnegie Mellon and the University of Texas at Arlington says it has a better way to train them. (arxiv.org) The paper, posted to arXiv on April 14, 2026, describes a system called Humanoid Transformer with Touch Dreaming, or HTD, for “contact-rich” tasks such as folding towels, organizing books, scooping cat litter, serving tea and inserting a T-shaped part. (arxiv.org) Before the robot can learn those tasks, it needs a body controller that keeps its legs and torso stable while its arms and hands work. The authors say they built that lower-body controller with reinforcement learning, a trial-and-error training method commonly used in robotics. (arxiv.org) They then used virtual-reality teleoperation to collect demonstrations from people, mapping human motion onto the humanoid so the robot could practice whole-body behavior instead of just hand motions at a fixed table. (arxiv.org) The “touch dreaming” part is a prediction step: alongside camera views and joint readings, the model is trained to anticipate future hand forces and compact tactile signals, so it can prepare for contact before it slips or jams. (arxiv.org) That setup targets a persistent robotics problem. Many manipulation systems work best when the robot stands still in a controlled workspace, but household and warehouse tasks often require walking, balancing, reaching and feeling contact at the same time. (humanoid-touch-dream.github.io) Across the five tasks in the paper, the authors report a 90.9% relative improvement in average success rate over their stronger baseline. They also report that predicting tactile information in a compressed latent form, rather than raw tactile values, produced a 30% relative gain in success rate. (arxiv.org) The project page includes real-world videos of the policy running those tasks on hardware, including towel folding, tea serving and a tight 3.5-millimeter insert task. One note on the site says several videos were recorded while the robot’s left hand had intermittent communication failures shortly before it went fully offline. (humanoid-touch-dream.github.io) The author list spans Carnegie Mellon University, UT Arlington and the Bosch Center for Artificial Intelligence. A public GitHub repository says code for the whole-body controller, teleoperation system and HTD policy training is planned for release by early May 2026. (humanoid-touch-dream.github.io) (github.com) The paper has not yet gone through peer review, and its results are limited to the five tasks and hardware setup shown by the authors. But it lands as more humanoid labs try to move from staged demos toward robots that can keep balance, use both hands and handle messy contact in the real world. (arxiv.org) (humanoid-touch-dream.github.io)