Source note
DEX-Mouse: A Low-cost Portable and Universal Interface with Force Feedback for Data Collection of Dexterous Robotic Hands
Summary
DEX-Mouse is a low-cost hand-held interface for collecting dexterous robot-hand demonstrations on real hardware. It aims to make data collection portable, calibration-free across operators, and more physically aligned with the target robot by adding force feedback and an attached forearm-mounted setup.
Problem
- Dexterous robot learning needs large amounts of real demonstration data that match the robot's actual kinematics and contact dynamics.
- Simulation and video pipelines lose fidelity through sim-to-real error, occlusion, and human-to-robot retargeting mismatch.
- Existing teleoperation devices such as MoCap gloves and other hand-held systems often need per-user calibration, custom hardware, fixed lab setups, or do not support multiple robot hands well.
Approach
- The paper builds a portable teleoperation device from off-the-shelf parts for under USD 150, with no per-operator calibration and no structural modification for different users.
- The interface uses tendon-driven fingers, a direct-driven thumb, and current-based kinesthetic force feedback so the user feels resistance when the robot hand is blocked by contact.
- It supports an attached configuration where the robot hand is mounted on the operator's forearm, plus a standard spatially separated teleoperation mode. The attached mode is meant to collect robot-aligned data with less coordinate mismatch.
- Hand motion is mapped to the robot with simple proportional retargeting instead of complex morphology-dependent retargeting. The firmware runs at 100 Hz; a VIVE tracker provides global pose, and an onboard camera records aligned RGB at 30 Hz.
Results
- In a user study with 8 participants across 3 interfaces and 2 collection configurations, DEX-Mouse in the attached setup reached 86.67% overall success with 10.05 s average completion time.
- Under the same attached setup, DEX-Mouse beat DOGlove and Manus glove on overall average success: 86.67% vs 77.5% and 62.5%. It also had the fastest attached completion time: 10.05 s vs 11.67 s and 11.93 s.
- Task breakdown for DEX-Mouse attached: 95.0% success / 5.57 s on pick-and-place, 72.5% / 14.29 s on peg-in-hole, and 92.5% / 10.29 s on hammering.
- DEX-Mouse teleoperation was worse than its attached mode: 52.5% overall success and 18.77 s average time, compared with 86.67% and 10.05 s when attached.
- Across all interfaces, attached collection outperformed remote teleoperation on overall average success and speed: 75.56% and 11.22 s vs 46.39% and 19.42 s.
- The paper also claims the attached configuration lowered perceived operator workload across compared interfaces, but the excerpt does not include the full workload table or exact questionnaire values.