Robots Atlas>ROBOTS ATLAS
Training

RFI

2026ResearchPublished: 29 September 2026Updated: 29 September 2026Published
Key innovation
Adds contact reactivity to an already pretrained VLA policy at the post-training stage by grafting a parallel force-driven action expert initialized as a zero residual, preserving the base model's general manipulation knowledge.
Category
Training
Abstraction level
Pattern
Operation level
Post-trainingTrainingRobot control
Use cases
Contact-rich manipulationCloth / towel foldingBook insertion onto a shelfRing placement (Tower of Hanoi)Post-training of vision-driven VLA policiesRobot control with a 6D force sensor

How it works

RFI is implemented through three architectural choices. (1) Dual action experts: a reactive action expert is grafted beside the original action expert; it decodes actions causally within an action chunk and consumes the most recent 6D end-effector force. (2) Causal force memory: force enters via a causal force memory and zero-initialized cross attention, so force history is incorporated without modifying the initial action output; the memory is latency-aligned with the robot state. (3) Within-chunk refresh: the slow vision-language prefix is computed once and cached, and each within-chunk refresh re-encodes the latest latency-aligned force history without another full forward pass through the vision-language module. Prior preservation comes from initialization: the reactive expert copies the original action weights and uses shifted causal attention so force is added as a zero-initialized residual, meaning that before post-training the force residual is exactly zero and the augmented model starts from the same action output as the base VLA. Training uses an additive flow-matching objective that jointly supervises both streams: vision-only batches mask the force memory while online correction batches activate it (a fixed 1:1 offline:online sampling ratio). A repeated online DAgger loop collects force-enabled human corrections (through Flexiv TDK) and mixes them with offline data to address policy-dependent distribution shift.

Problem solved

Pretrained VLA policies are largely vision-driven and fail in contact-rich manipulation when the scene is occluded, depth is ambiguous, and small force errors push execution off the offline demonstration distribution. Force reveals contact states vision alone misses (whether cloth was grasped, a book bottomed out, or a ring jammed). RFI adds this reactivity without discarding the base model's knowledge.

Components

Reactive Action Expert (Dual Action Experts)Force-driven reactive action stream

A second action expert grafted beside the original action expert of the base VLA. It decodes actions causally within an action chunk and consumes the most recent 6D end-effector force, forming a parallel reactive stream without replacing the base policy.

Causal Force Memory (zero-initialized cross attention)Force history injection

The channel through which force signals enter the policy via zero-initialized cross attention. It incorporates force history without modifying the initial action output and provides a latency-aligned force memory synchronized with the actual robot state.

Zero-Initialized Residual / Prior PreservationPreserving base-model priors

An initialization mechanism: the reactive expert copies the original action weights and uses shifted causal attention so force is added as a zero-initialized residual. Before post-training the force residual is exactly zero, so the augmented model starts from the same action output as the base VLA.

Within-Chunk RefreshLow-latency reactivity

During execution the slow vision-language prefix is computed once and cached. Each within-chunk refresh re-encodes the latest latency-aligned force history without another full forward pass through the vision-language module.

Additive Flow-Matching Objective (Heterogeneous Data Training)Joint training objective

An additive flow-matching objective jointly supervises the base and reactive streams. Vision-only batches mask the force memory while online correction batches activate it, allowing one model to train on both dataset types without synthetic labels; a fixed 1:1 offline:online sampling ratio is used.

Online DAgger LoopDistribution-shift correction

A repeated online DAgger loop collects force-enabled human corrections (through Flexiv TDK) and mixes them with offline data to address policy-dependent distribution shift.

Official

Implementation

Implementation pitfalls
Losing base-model priors when adding forceHigh

Naively adding a force pathway can disturb the pretrained VLA output and erase its general manipulation knowledge.

Fix:Initialize force as a zero residual (copy expert weights + shifted causal attention) so the model starts from the same output as the base VLA.
Force memory latency misalignmentHigh

If force history is not synchronized with the actual robot state, reactive actions respond to stale contact.

Fix:Use latency-aligned force memory and a within-chunk refresh that re-encodes the latest force history.
Policy-dependent distribution shiftMedium

Small force errors push execution off the offline demonstration distribution, which offline data alone does not cover.

Fix:A repeated online DAgger loop collecting force-enabled human corrections and mixing them with offline data (1:1 ratio).

Evolution

Original paper · 2026 · CoRL 2026 · Yi Wang
Never Too Late for Force: Accelerating VLA Post-Training with Reactive Force Injection
Yi Wang, Wendi Chen, Zimo Wen, Han Xue, Xueqi Li, Wenye Yu, Zhijie Chen, Hao Yang, Jun Lv, Chuan Wen, Cewu Lu
2026
RFI introduced in the LIFT method
Inflection point

The paper "Never Too Late for Force" (arXiv:2607.14236, CoRL 2026) introduces Reactive Force Injection as the core of the LIFT framework for force-aware VLA post-training.

Hyperparameters (configurable axes)

Offline:online sampling ratioHigh

Mixing proportion of offline data and online corrections during post-training. The paper uses a fixed 1:1 value.

1:1Value used in the paper; outperformed online-only training on towel folding.
Force memory length (latency-aligned)High

Span of 6D force history fed to the reactive expert through the causal force memory; must be latency-aligned with the robot state.

Action chunk sizeMedium

Number of actions decoded per chunk; determines how often the within-chunk refresh with the latest force history occurs.

Execution paradigm

Activation pattern
Input dependent
Routing mechanism

Parallelism

Parallelism level
Partially parallel

The slow vision-language prefix is computed once and cached, while fast within-chunk force refreshes run without a full vision-language forward pass.

Scope
Inference