"Touch Dreaming" Gives Humanoid Robots 90.9% Better Dexterity — CMU & Bosch Paper
The Research
Researchers from Carnegie Mellon University and the Bosch Center for AI have developed a humanoid robot control system that "dreams" — predicts — upcoming tactile signals rather than merely reacting to them. The Humanoid Transformer with Touch Dreaming (HTD) achieved a 90.9% improvement in average success rate over the stronger vision-only baseline across five dexterous real-world tasks.
How It Works
HTD combines imitation learning with prediction of future contact-related signals. Instead of reacting to current tactile feedback, the robot anticipates how touch and force will evolve moments ahead and coordinates movement accordingly. The architecture integrates whole-body reinforcement learning, upper-body inverse kinematics, and dexterous hand retargeting.
Tasks Tested
- T-shape part insertion
- Book organization
- Towel folding
- Cat litter scooping
- Tea serving
Why It Matters
Contact-rich dexterous manipulation remains one of the hardest open problems in humanoid robotics. HTD's open-source release means other researchers and robot manufacturers can immediately test the approach. If the gains transfer to commercial platforms, the work accelerates deployment in manufacturing, logistics, and domestic service.
Source: TechXplore | arXiv paper
Related Articles
Figure’s robot fleet has crossed its human headcount for the first time, according to CEO Brett Adcock. The chart shared with the tweet points to a steep 2026 production ramp and a workforce line near the mid-hundreds.
A new MIT Technology Review investigation reveals that humanoid robot companies routinely obscure the scale of human teleoperation and data collection labor behind their demos, risking a repeat of AI's early automation-washing scandals.
The Allen Institute for AI released MolmoAct 2 on May 5, a fully open-source 7B-parameter robot foundation model that surpasses Physical Intelligence's π0.5 across simulation and real-world tasks, achieving up to 87.1% success on zero-shot manipulation.