Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

DexRemo: A Force-Sensing Hand Platform and Task Suite for Contact-Rich Removal Manipulation

Liman Wang, Hanyang Zhong, Sun Zhaole, Wenting Cao, Babar Jamil, Andy Tyrrell, Shan Luo, Jihong Zhu

November 2026

TL;DR A low-cost three-finger hand with three-axis force sensors in every fingertip, plus a nine-task real-robot suite for peeling, scraping, wiping and digging targets free. Giving an imitation policy fingertip poses and forces rotated into a shared wrist frame raises success on six core tasks to 83%, against 55% with the same forces in each fingertip's local frame.

What they do

Peeling a glued part off a build plate, scraping off Blu-Tack, wiping a stain or digging a fossil out of sand all depend on contact states that cameras and joint encoders barely see. DexRemo is a low-cost three-finger hand with force sensors in every fingertip, a nine-scenario real-robot task suite for this kind of removal manipulation, and a controlled study of how to represent multi-finger forces for imitation learning.

How it works

  • Hand. Three identical 4-DoF fingers driven by 12 servos, 1.8 kg, with interchangeable rigid wedge or soft cleaning fingertips. Each fingertip holds three load cells in a triangle, calibrated to R² > 0.986 on every axis. The hand costs about $200, plus $170 per sensing fingertip.
  • Wrist-frame forces. Forward kinematics rotates each fingertip force from its sensor frame into a shared wrist frame, so a force component points the same way for every finger and posture.
  • Policy. ACT takes three RGB views, proprioception, fingertip poses and forces as separate tokens and predicts 100-step joint-displacement chunks, executing 50 before replanning at 15 Hz. Each task gets 20 demonstrations, teleoperated with a passive replica of the hand.
  • Suite. Nine tasks across adhesive detachment and fingertip self-cleaning, surface mark removal, and retrieval from sand or cat litter, with preparation procedures, physical success criteria and held-out objects, colors, poses or attachments.

Results

  • Pullout strength is 62.5 N, the highest among the hands compared (LEAP Hand 19.5 N, Allegro 8.5 N).
  • Six core tasks, three seeds with 30 rollouts each: success is 24% without force or fingertip pose, 55% with pose and local-frame forces, and 83% with pose and wrist-frame forces. Wrist-frame forces beat local-frame forces on all six tasks.
  • On the three held-out extension tasks the full policy succeeds 8/10, 9/10 and 7/10. In a single-seed check, ACT beats Diffusion Policy on three tasks, e.g. 28/30 vs 6/30 on removing a printed part from the build plate.
  • Replacing the camera images with black frames drops both tested tasks to 0/30.

Limitations

Only local tabletop removal: the wrist pose is fixed and a linear stage provides vertical motion, with no arm or mobile base. Material preparation still varies between trials. Demonstrations come from teleoperation without haptic feedback, policies are trained offline, and the ACT versus Diffusion Policy comparison uses one seed.

Figures

From Figure 1: (a) The DexRemo hand removes 3D-printed barnacles glued with hot-melt adhesive to a turtle-shell mocku...
From Figure 1: (a) The DexRemo hand removes 3D-printed barnacles glued with hot-melt adhesive to a turtle-shell mockup. (b) Human removal behaviors. (c) Fingertip forces are rotated from each sensor frame into a shared wrist frame and fused with vision, fingertip poses and proprioception; an ACT policy outputs action chunks that debond the target from its edge.
From Figure 2: Left: sequences from the nine tasks in three families, adhesive removal (A1-A5), surface cleaning (B1-...
From Figure 2: Left: sequences from the nine tasks in three families, adhesive removal (A1-A5), surface cleaning (B1-B2) and buried retrieval (C1-C2). Top right: joint states, wrist-frame fingertip forces, fingertip poses and RGB images are encoded and fused by an Action Chunking Transformer that predicts short joint-displacement chunks. Bottom right: the leader-follower teleoperation setup, with several RGB cameras and force-sensing fingertips on the follower hand.
From Figure 3: Hand design. (a) Kinematic layout of the three fingers. (b) Each finger has two flexion joints (IP, MC...
From Figure 3: Hand design. (a) Kinematic layout of the three fingers. (b) Each finger has two flexion joints (IP, MCP) and a two-axis base joint (CMC). (c) The physical prototype with three identical thumb-like fingers.
From Figure 8: Success rates on the six core tasks for each input configuration, all with RGB and proprioception (R: ...
From Figure 8: Success rates on the six core tasks for each input configuration, all with RGB and proprioception (R: local-frame forces, P: fingertip poses, W: wrist-frame forces). Points are means over three seeds with 30 real-robot rollouts each; P+W is highest on A1-A4, B1 and C1.
Abstract

Many manipulation tasks require separating adhered, embedded, contaminated, or mechanically constrained targets. We study this setting as removal manipulation, where progress depends on evolving contact states that are only partially observable from vision and kinematics. We introduce DexRemo, a three-finger robotic hand with high-force actuation, modular fingertips, and three-axis fingertip force sensing, together with a nine-scenario real-robot task suite spanning adhesive detachment, surface cleaning and self-cleaning, and occluded retrieval. To make multi-finger force observations easier to fuse, we express fingertip forces in a shared wrist-centered frame, giving force components consistent geometric meaning across fingers and time. Using an Action Chunking Transformer backbone, policies fuse multi-view vision, proprioception, fingertip pose, and force observations to predict closed-loop action chunks from demonstrations. Evaluation uses six core benchmark tasks for controlled fingertip-input and force-frame ablations, plus three extension scenarios for additional held-out coverage. Across these evaluations, the learned policies achieve consistent removal performance, and ablations show that wrist-frame forces, especially with fingertip pose, improve stability and generalization over local-frame force inputs and no-force baselines. Project website: https://dexremo.github.io

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.