TouchBridge trains dexterous robots from human-only demos, hitting 72.5% success rate
ZeYanjie · x · 2026-11-05
TouchBridge is a hardware-software co-designed framework that converts human demonstrations into robot training data. Its Tac-Retarget method aligns fingertip contact patches via tactile sensing, requiring no object meshes, pose estimates, or physics simulation.
Key results (4 real-world contact-rich tasks like page turning and light-bulb screwing):
- Policies trained on human-only data achieve 72.5% average success, close to 77.5% with an equal number of teleoperated robot demos
- Human demo collection is 7.0× faster
- Pipeline: wearable capture interface + tactile-conditioned retargeting + visual synthesis + camera-anchored action conversion to fine-tune a pretrained VLA model
Paper on arXiv; website includes demos.
More from Embodied
- Unitree's UMI Humanoid Gains New Capabilities With F/T Sensors — HaozhiQ · 2026-11-05
- Force sensors are the new eyes guiding humanoid policies, researchers say — HaozhiQ · 2026-11-05
- FLoRe Teaches Humanoids How Much Force to Apply From Robot-Free Human Demos — HaozhiQ · 2026-11-05
- Perceptron Egocentric 1.0 turns egocentric video into robot training data at $29/hour — rohanpaul_ai · 2026-11-05
- Curbie Robot Autonomously Hauls Trash Bins to the Curb Through Snow and Ice — heatherknight · 2026-11-05
- FUSE physical AI localization conference lands in SF Nov 19 with Anduril CEO keynote — kscottz · 2026-11-05