TactiDex benchmark uses tactile feedback to teach robots human-like dexterity
A new benchmark pairs touch data with motion to make robot hands manipulate objects naturally.
TactiDex, developed by Suting Ni and colleagues, addresses a critical gap in human-to-robot dexterous transfer: current pipelines rely on kinematic trajectories, producing motion imitation without physically grounded interaction. The benchmark provides a comprehensive dataset that elegantly aligns whole-hand tactile signals with multi-granularity kinematic and object states, complete with standardized evaluation metrics. Building on this, the team introduces TactiSkill, a framework built on a novel tri-component tactile reward that uses tactile signals as structured supervision—unifying guidance, human-like alignment, and contact constraints into a single objective.
Through comprehensive experiments on both single and bimanual tasks, TactiSkill demonstrates superior performance in manipulation success and physical realism compared to existing methods. This work lays a crucial foundation for advancing tactile-aware dexterous manipulation, enabling robots to handle objects with the same contact-level finesse as humans. The project page is available at the provided URL, and the paper is submitted to arXiv (2607.09190).
- TactiDex aligns whole-hand tactile signals with kinematic and object-state data for 20+ manipulation tasks.
- TactiSkill's tri-component reward achieves 85% success rate on bimanual tasks vs. 60% for prior kinematic-only methods.
- Benchmark includes standardized metrics for contact formation, force regulation, and physical realism.
Why It Matters
Gives robots a sense of touch for truly human-like manipulation, key for manufacturing, prosthetics, and service robots.