LIFT: Never Too Late for Force

Abstract

Pretrained vision-language-action (VLA) policies provide strong language-conditioned manipulation knowledge, but they remain largely vision-driven and can struggle once manipulation enters contact states where the scene is occluded, depth is ambiguous, or small force errors push execution off the offline demonstration distribution.

LIFT (Late Reactive Injection of Force for VLA Post-Training) is a force-aware post-training framework that adds contact reactivity while preserving the policy's output at initialization. LIFT grafts a reactive action expert beside the original action expert, initializes it from pretrained action weights, and injects recent 6D end-effector force through causal force memory and zero-initialized cross attention, enabling actions to be refreshed during execution. To address the policy-dependent distribution shift of contact feedback, LIFT uses an online DAgger loop with a mixture of offline task-alignment data and human-corrected online rollouts.

Across towel folding, book insertion, and Hanoi ring placement, LIFT learns faster and reaches higher peak and final performance than vision-only DAgger, while ablations show that reactive force memory and online corrective data are both important for robust contact-rich manipulation. Our code is publicly available.

Method

Force reveals contact states that a wrist camera can miss: whether cloth was grasped, a book has bottomed out, or a ring is jammed. LIFT introduces this feedback during post-training, building on a pretrained VLA and the vision-only demonstrations already available for a task.

Three objectives guide the design: react to new contact, preserve the pretrained action output at initialization, and learn jointly from visual demonstrations and force-enabled corrections.

Closing the loop with reactive force injection O1

LIFT starts from a pretrained vision-only VLA and grafts a reactive action expert beside the original action expert. Instead of committing to a full open-loop action chunk, the reactive stream decodes actions causally within the chunk and receives recent 6D end-effector force as latency-aligned force memory.

At runtime, the slow vision-language prefix is computed once and cached. Each within-chunk refresh re-encodes the latest latency-aligned force history and reevaluates both action experts against the cached prefix. The reactive expert supplies the full action output, allowing contact updates without another full vision-language forward pass.

**LIFT architecture.** Causal force memory enters the reactive expert through zero-initialized cross attention, enabling within-chunk updates while preserving the policy output at initialization.

Preserving the pretrained VLA prior at initialization O2

Shifted causal attention mask — **Shifted causal attention.** Each reactive token sees the vision-language prefix, later base-action tokens, and the causal reactive prefix.

Adding a force pathway should not erase the pretrained VLA prior. LIFT copies the original action-expert weights into the reactive expert and uses a shifted causal attention pattern so each reactive token receives an equivalent context to the original fully attentive action token at initialization.

The force-injected cross attention is added as a residual update with a zero-initialized output projection. Before post-training, the force residual is exactly zero, so the augmented model starts from the same action output as the base VLA.

Training with heterogeneous visual and force data O3

Train both action streams. An additive flow-matching objective jointly supervises the base and reactive streams on both data sources. At inference, both streams are computed because shifted causal attention connects them, while only the reactive action is sent to the robot.

Enable force only when measured. Vision-only batches mask encoded force memory before cross attention, blocking gradients into the force pathway. Online correction batches activate measured force memory, so one model can train on both datasets without synthetic force labels.

Keep both supervision sources. Each update uses a fixed 1:1 offline:online sampling ratio: demonstrations provide task coverage, and online corrections cover states visited by the current policy.

From task alignment to online correction

LIFT two-stage training pipeline — **Two-stage training.** Align on vision-only demonstrations, then close the loop between force-enabled robot corrections and online post-training.

1. Learn from visual demonstrations

Train on handheld demonstrations with force masked out. This stage aligns the pretrained policy to the target task using the abundant vision-only dataset.

2. Adapt to contact on the robot

Collect force-enabled human corrections through Flexiv TDK, mix them with the offline dataset, and redeploy updated checkpoints. Repeated online aggregation tracks contact states induced by the changing policy.

Experimental results

We evaluate towel folding, book insertion, and Hanoi ring placement on a Flexiv Rizon 4S with a 6D end-effector force sensor. Vision-only demonstrations come from a handheld device; online force-enabled corrections are collected with Flexiv TDK.

Evaluation. Main learning curves and ablations use three groups of ten autonomous rollouts per checkpoint (n = 30), with the reported 95% confidence intervals. Towel and book use graded task scores; Hanoi uses binary success. Shift tests retain 10 rollouts per condition.

Training budget. Policies are compared at similar training steps, such as 1,000 and 2,000, within a comparable time budget. Each online experiment lasts approximately 2–3 hours and collects roughly 20–30 episodes.

Six policies in the comparison

LIFT: The full reactive force-injected VLA trained with online DAgger.
π_0.5 w/ Online DAgger: The same online loop without force input.
LIFT w/o Reactive Force Injection: Force post-training with a single-frame force input instead of reactive force memory.
LIFT w/ Offline DAgger: The reactive model trained with a fixed force-correction buffer, without repeated on-policy aggregation.
π_0.5 w/ Offline Handheld Data: Vision-only training on the larger offline handheld dataset.
Residual Policy: A lightweight force-reactive Transformer that adds action corrections to a frozen, offline-trained π_0.5 policy.

Does force accelerate or improve post-training? Q1

Force-enabled post-training accelerates learning and raises peak and final performance over vision-only online DAgger across the three tasks.

Task scores over training steps for towel folding, book insertion, and Hanoi ring placement, with online and offline baselines — **Learning curves on three tasks.** Force-enabled post-training improves adaptation over vision-only online DAgger. Each checkpoint uses n = 30 rollouts with 95% confidence intervals.

Towel folding. Force exposes whether the thin cloth was actually grasped, allowing recovery from empty grasps that are difficult to detect with a monocular wrist camera.

Book insertion. Contact feedback helps identify when to push, stop, or release inside the constrained shelf slot. Vision-only control can keep pushing after the book bottoms out or release before it is seated.

Hanoi ring placement. The small ring-pole clearance makes slight misalignment likely to cause a jam. Force direction and magnitude help LIFT identify offset or stuck states and correct its motion.

Does reactivity matter? Q2

Comparing LIFT with the single-frame force variant in the main learning curves shows why recent contact history and timely updates matter. The benefit is strongest when the contact state evolves throughout execution.

Book: infer the contact phase

The same current load can mean insertion is still in progress or the book has already bottomed out. Force memory helps identify the phase and trigger release or withdrawal instead of continued pushing.

Hanoi: distinguish impacts from jams

A transient impact can destabilize single-frame control. Recent force history and fast action updates help separate an initial tap from persistent misalignment and correct the placement.

Towel: an immediate cue can suffice

A single force frame can already reveal whether cloth is grasped, so the single-frame variant can match or outperform LIFT. Both remain stronger than the vision-only baseline.

Does LIFT preserve original VLA generalization? Q3

The final LIFT checkpoint shows no clear performance degradation relative to its in-distribution behavior under the tested object, tablecloth, and lighting changes. These evaluations retain 10 rollouts per condition.

Final LIFT checkpoint performance under original conditions and tested object, tablecloth, and purple-light changes — **Generalization of the final LIFT checkpoint.** No clear degradation under the tested shifts; ten rollouts per condition.

View the generalization settings

Photographs of the generalization settings — A blue towel or gray book tests object changes; an alternate tablecloth and purple side light test environmental changes. The robot start pose and target location stay fixed.

Does online data matter for reactive force injection? Q4

LIFT w/ Offline DAgger uses a fixed force-correction buffer and underperforms on all three tasks, scoring zero on book insertion. These offline-only results appear as separate markers in the main learning curves. Repeated online DAgger exposes the policy to failures induced by its own behavior and supplies corrections for the resulting contact distribution.

Why use LIFT rather than a lightweight residual policy? Q5

The lightweight residual policy achieves substantially lower scores than LIFT under the tested base-policy and intervention protocol. The weak base requires large corrections, while sparse residual targets make intervention timing difficult to learn. LIFT instead learns full actions from demonstrations and corrections; implementation details are in the paper's appendix.

Learning curves of the lightweight residual policy on the three tasks — **Lightweight residual baseline on three tasks.** The residual module underperforms LIFT under the tested base-policy and intervention protocol.

Why retain offline data during online post-training? Q6

Offline demonstrations broaden behavioral coverage and help preserve the pretrained prior. On towel folding, online-only training (0:1) performs substantially worse than 1:1 and 1:2, with 1:1 improving faster. All main experiments use fixed 1:1 offline:online sampling.

Towel folding scores for offline-to-online sampling ratios of 1:1, 0:1, and 1:2 — **Offline data ratio ablation.** Mixed training outperforms online-only training; n = 30 with 95% confidence intervals.

Task protocols and failure cases

Each rollout lasts at most two minutes. These examples illustrate successful behavior and common failures across the three tasks.

Qualitative examples of towel folding, book insertion, and Hanoi ring placement, with checks and crosses marking successful and failed cases — **Task protocols and failure cases.** Individual examples from the paper's appendix, with success and failure marked by checks and crosses.

Scoring rules and common failures

Towel folding: The highest completed stage determines the score: first grasp 0.25, first fold 0.5, second grasp 0.75, and second fold 1.0. Common failures include empty grasps, the wrong second corner, and incorrect folding motions.
Book insertion: Grasping, lifting, and inserting without collision earns 0.5. Closing the gripper, pushing the spine to complete insertion, and disengaging earns 1.0. Failures include side collisions, early release, over-pushing, missing the spine, and incomplete insertion or disengagement.
Hanoi ring placement: A ring fully seated on the pole base earns 1; otherwise the score is 0. Common failures include angled insertion and jamming, or impact-induced oscillation and drift.

Task videos

Play a clip to start the matching demonstrations together.

Towel Folding

LIFT

π_0.5 w/ Online DAgger

Book Insertion

LIFT

π_0.5 w/ Online DAgger

LIFT w/o Reactive Force Injection

Hanoi Ring Placement

LIFT

π_0.5 w/ Online DAgger

LIFT w/o Reactive Force Injection

Conclusion and limitations

LIFT adds force-aware post-training to pretrained VLAs through reactive force injection, prior-preserving initialization, and online corrective training. Across the three tasks, force feedback improves learning and peak and final performance, while the final policy shows no clear degradation under the tested object and environmental changes.

Limitations and future directions. Human corrections limit data throughput, and the evaluation covers one robot arm. Future work should reduce the correction load and test more robot arms, force sensors, and end-effectors.

BibTeX

@article{wang2026never,
  title={Never Too Late for Force: Accelerating VLA Post-Training with Reactive Force Injection},
  author={Wang, Yi and Chen, Wendi and Wen, Zimo and Xue, Han and Li, Xueqi and Yu, Wenye and Chen, Zhijie and Yang, Hao and Lv, Jun and Wen, Chuan and others},
  journal={arXiv preprint arXiv:2607.14236},
  year={2026}
}