ReForce — force-aware retargeting via closed-loop residual force tracker

TL;DR

A 5-tweet thread from @wu876768 announcing ReForce, a closed-loop force tracker that sits between an upstream motion+force reference (from teleop or a learned policy such as ACT) and a dexterous robot hand, predicting residual joint commands from motion/force targets plus live tactile feedback. The framing argument is that human data can scale robot learning and geometric retargeting can transfer motion, but manipulation is decided by force — so retargeting without a force channel silently fails on deformables and precise-contact tasks (missed contact, slip, crush). The reported real-robot delta on paper-cup grasping is force-safe success 30%→70% with severe missing-contact failures dropping 7/10→0/10 when the reference is a learned ACT policy. Demos span spring pinching, sequential finger release, board wiping, tongs squeezing, and cup pouring.

Key claims

Geometric-only retargeting from human to robot loses the force channel, and the same motion trace can miss contact, slip, or crush a deformable — force must be a first-class transferred signal, not a post-hoc controller correction [1/5].
ReForce learns its residual force tracker from simulated interaction rollouts, then deploys the same tracker on real hardware with live tactile feedback as input [2/5].
At deployment the tracker consumes a motion+force target plus current tactile state and emits a residual joint command that composes with the upstream reference [2/5].
On paper-cup grasping ReForce achieves the lowest force-tracking error against baselines in both settings tested (open-loop reference and learned-policy reference) [3/5].
With a learned ACT policy as the reference, force-safe success climbs from 30% → 70% and severe missing-contact failures drop from 7/10 → 0/10; contact engagement is reported as the strongest of the compared configurations [3/5].
The system runs from two upstream reference types — human teleoperation and learned ACT policies — with the same residual force tracker, on demos spanning spring pinching, sequential finger release, board wiping, tongs squeezing, and cup pouring [4/5].

Method

The thread’s method sketch: at collection time, an upstream reference produces a paired motion + force trajectory. ReForce trains a closed-loop force tracker in simulation whose input is the current motion/force target plus live tactile feedback, and whose output is a residual joint command added to the reference. Because training is in sim, force rollouts are cheap, and the tracker is agnostic to whether the reference is teleop or a learned policy at deployment. The thread does not disclose the tactile-sensor abstraction, the simulator, the residual-command parameterization, or the base robot hand — those would need a paper release to verify. No linked paper, project page, or code URL was included in the thread.

Results

Only one quantitative task is reported in the thread — paper-cup grasping — under two upstream reference settings:

Lowest force-tracking error vs unspecified baselines in both settings [3/5].
Learned-policy reference (ACT): force-safe success 30% → 70%, severe missing-contact failures 7/10 → 0/10; contact engagement described as strongest [3/5].
Real-robot demo tasks (qualitative, no metrics): spring pinching, sequential finger release, board wiping, tongs squeezing, cup pouring [4/5].

No comparison to specific prior force-aware methods, sensor spec, robot spec, or training-scale numbers are given in the thread.

Why it’s interesting

Sits at the intersection of two open questions already tracked in the wiki. On Human-to-Robot Retargeting, ReForce stakes out a distinct position on the interface-location axis that no filed paper occupies: the shared human–robot interface is a force-tracking residual, learned in sim and applied at deployment, rather than a reduced action space (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), a physics-in-loop bilevel retargeter (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting), a kinematic 4D-HOI reconstruction (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), or an object-centric contact-wrench transfer (CHORD: Object-Centric Contact Wrench Guidance for Dexterous Manipulation). On Tactile sensing for manipulation, it is the policy-composition counterpart to FACT: Force-Aware Contact-Rich Manipulation via Timestep Modulation — demystifying when and why VLAs fail on contact-rich tasks FACT — FACT modulates a VLA per-layer by current force, ReForce composes a residual atop an existing motion+force reference — both argue force must shape the policy rather than be appended as an input. The 30%→70% force-safe success delta on paper-cup grasping is the specific gap FACT’s authors and Tactile Genesis’s authors have both flagged as under-served by pure-motion retargeting.