HTTP/2 301 server: GitHub.com content-type: text/html location: https://clairelc.github.io/dexforce.github.io/ x-github-request-id: 0FF1:3ABDEF:7F8099:8F37F5:6951A6C2 accept-ranges: bytes age: 0 date: Sun, 28 Dec 2025 21:53:08 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210057-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1766958788.403061,VS0,VE199 vary: Accept-Encoding x-fastly-request-id: 5db01384d9d89f6872b6cd2fcefc93cabdb81172 content-length: 162 HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Wed, 21 May 2025 19:29:06 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"682e2982-f79b" expires: Sun, 28 Dec 2025 22:03:08 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: A137:318CF6:7F6F25:8F233D:6951A6C4 accept-ranges: bytes age: 0 date: Sun, 28 Dec 2025 21:53:08 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210057-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1766958789.615574,VS0,VE220 vary: Accept-Encoding x-fastly-request-id: 78d4a32b6eb1ec86fd155e1e1edf0425de8ba446 content-length: 5755

DexForce

Extracting Force-informed Actions from Kinesthetic Demonstrations

for Dexterous Manipulation

Claire Chen, Zhongchun Yu, Hojung Choi, Mark Cutkosky, Jeannette Bohg

Stanford University

RA-L June 2025

  Read the paper   Watch a 5 minute summary (captioned)   BibTex

DexForce leverages contact forces, measured during kinesthetic demonstrations, to compute force-informed actions.

We use DexForce demonstrations to learn policies for6 tasksand evaluate each rollout according to the following criteria:

slide cube

success

failure

reorient smiley

success

partial success

failure

open AirPods

success

partial success

failure

grasp battery

success

partial success

failure

unscrew nut

success

partial success

failure

flip box

success

partial success

failure

We show that DexForce's force-informed actions are critical for learning capable policies.

Policies trained on our force-informed actions achieve an average76%success rate across all 6 tasks.
Baseline policies trained on actions thatdo notaccount for contact forces havenear zerosuccess rates.

To view policy rollouts trainedwithandwithoutforce-informed actions for each task, view on larger screen.

Explore policies trainedwithandwithoutforce-informed actions for each task:

slide cube reorient smiley open AirPods grasp battery unscrew nut flip box

with force-informed actions

without force-informed actions

with force-informed actions

without force-informed actions

with force-informed actions

without force-informed actions

with force-informed actions

without force-informed actions

with force-informed actions

without force-informed actions

with force-informed actions

without force-informed actions

Hardware

We instrument an Allegro hand with a wrist-mounted Intel RealSense D435 camera (+ fisheye lens attachment) and two CoinFT six-axis force-torque sensors.

To re-create our setup:
Fisheye RealSense: We have open-sourced our camera and fisheye lens mounting parts. Get the instructions and partshereand pleasecite usif you find these useful!
CoinFT: The developers of CoinFT plan to release the sensor design soon!

Supplementary video