TouchBridge uses touch to transfer human demonstrations
TouchBridge: Tactile-Conditioned Human-to-Robot Transfer for Dexterous Manipulation · arXiv:2610.40349
In plain words
TouchBridge is a system for turning demonstrations by people into training examples for a dexterous robot. It records what the person sees, how their hand moves and where their fingertips touch, then uses that touch information to transfer the demonstration to the robot.
On four real-robot tasks, policies trained on TouchBridge-converted demonstrations averaged 72.5% success, compared with 77.5% for policies trained on the same number of robot demonstrations. The result applies to those tasks and trials.
The research
When a person demonstrates a task that depends on steady fingertip contact, how can a robot learn the interaction rather than copy only the hand’s movement? The question matters because a transferred motion can look plausible while losing the contact that makes the task succeed.
TouchBridge uses fingertip touch measurements to convert human demonstrations into robot training data while preserving contact across different hands.
TouchBridge’s Tac-Retarget method uses fingertip measurements to carry demonstrated contacts across human and robot hands.
Researchers compared policies trained on its converted data with policies trained on data converted by DexPilot and UniDex, using the same 60 human demonstrations per task and identical training settings.
Across four real-robot tasks, Tac-Retarget policies averaged 72.5% success, exceeding both alternatives. In a separate Grasp Egg Roll simulation diagnostic, transferred contact locations were closer to the human touch centers than with either alternative. Together, these results support contact-aware transfer as useful for the tested tasks.
These results cover four tasks on one real-robot platform, with 20 real-world trials per task for policy success. The contact-error diagnostic and its retargeting component ablations use only Grasp Egg Roll and 20 demonstrations replayed in simulation.
Separate real-robot policy ablations across the four tasks found lower average success when tactile components were removed; exact ablation averages are not reported. The authors describe the demonstration set as modest and identify larger, more diverse datasets as future work.
What the evidence shows
- Why motion matching falls short
TouchBridge is a hardware–software framework for turning human demonstrations of contact-rich hand tasks into robot training data. The problem is that matching a person's hand pose or fingertip position can still move a contacting finger away from its support, disrupt simultaneous touches or break contact.
The framework therefore uses measured fingertip touch to guide motion transfer, rather than relying on hand motion alone.
- How TouchBridge transfers contact
Tac-Retarget, TouchBridge’s motion-transfer method, aligns the touched regions on human and robot fingertips and preserves the spatial relationships among simultaneous contacts. TouchBridge then synthesizes robot-view observations and constructs actions for training.
In the four-task policy comparison, each retargeting method converted the same 60 human demonstrations per task, and the resulting GR00T N1.7 policies used identical training settings.
- Task success against alternatives
The researchers evaluated policy success on four real-robot tasks, using 20 trials per task; success means meeting that task’s success criterion. Policies trained on Tac-Retarget-converted data averaged 72.5% success, 25.0 percentage points above the DexPilot comparison and 42.5 points above UniDex. The mean is across the four task success rates, so it does not describe every task or trial.
- Evidence about contact preservation
A separate simulation diagnostic tested where transferred contacts landed, not whether a learned policy completed a task. On the same 20 Grasp Egg Roll demonstrations, Tac-Retarget had lower mean contact error than DexPilot and UniDex.
In that task’s component checks, removing contact alignment or the contact graph increased error; separate real-robot policy ablations also found lower average success when tactile components were removed.
- How human data compare
The human-data comparison provides a different check: policies trained on 60 converted human demonstrations per task averaged 72.5% success, versus 77.5% for policies trained on the same number of teleoperated robot demonstrations.
Separately, human demonstrations took less time to collect per successful example. These results support the framework in the tested setting, while the authors identify the modest demonstration set as a boundary and call for larger, more diverse data.
Deep
- Original paper authors
- Haozhe Zhang, Gu Zhang, Zirui Zou, Yuzheng Chen, Jinlong Wu
- Submitted to arXiv
- License
- arXiv non-exclusive