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 reaches higher performance than vision-only post-training, with task-dependent adaptation dynamics and force-history benefits. Our code is publicly available.

Method

Reactive Force Injection

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.

Prior-Preserving Initialization

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 Data

LIFT trains one model on both abundant vision-only task-alignment data and scarce force-enabled online corrections. Vision-only batches use zero force placeholders and mask the encoded force memory, while online correction batches keep measured force active.

The main post-training protocol uses a fixed 1:1 offline:online sampling ratio. Offline demonstrations provide nominal action supervision, while online DAgger continually adds force-enabled corrections from states visited by the current policy.

LIFT training pipeline — **Training pipeline.** Vision-only alignment followed by online force-correction post-training.

Experiments

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.

Horizontal axis. Curves track optimization steps. A recorded sample is one 10 Hz frame; data reuse and policy-dependent intervention rates mean equal training steps do not imply equal robot interaction or human-correction budgets.

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 28M-parameter force-reactive Transformer that adds action corrections to a frozen, offline-trained π_0.5 policy.

Does LIFT accelerate post-training? Q1

LIFT reaches stronger final performance than vision-only online DAgger, with different adaptation dynamics across tasks. Towel folding improves earlier; the advantage on book insertion emerges later. Extended Hanoi training largely plateaus, supporting the higher final level beyond the original endpoint.

Task scores over training steps for towel folding, book insertion, and Hanoi ring placement, including extended Hanoi training and offline baselines — **Learning curves on three tasks.** Double-lined Hanoi curves extend training beyond the original endpoint. Each checkpoint uses n = 30 rollouts with 95% confidence intervals.

Why adapt the full action expert? Q2

The 28M-parameter Residual Policy adds force-conditioned corrections to a frozen base; LIFT's 300M-parameter reactive expert predicts full actions from pretrained action weights. The smaller residual underperforms LIFT under the tested base-policy and intervention protocol.

LIFT receives full-action supervision from both demonstrations and corrections. The residual learns expert-minus-base corrections during interventions and zero targets on offline or non-intervention samples. This comparison changes capacity, initialization, and supervision together; it does not isolate model size or establish that residual policies fail with stronger bases.

Learning curves of the 28M residual policy on towel folding, book insertion, and Hanoi ring placement — **Lightweight residual baseline.** A custom force-reactive Transformer inspired by ImplicitRDP and CR-DAgger, evaluated with n = 30 and 95% confidence intervals. This is not a reproduction of either method.

When does force history help? Q3

The comparison with single-frame force input in the main learning curves shows that the value of force history depends on the contact behavior.

Towel folding

Force helps verify whether the towel was grasped. A single force frame can outperform history when this immediate contact cue is sufficient.

Book insertion

Recent force evolution helps distinguish insertion phases, including reaching the back of the shelf and deciding when to stop pushing or release.

Hanoi ring placement

Force memory helps distinguish a persistent placement error or jam from the transient impact of an initial tap against the pole.

Why retain offline data? Q4

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

Online corrections concentrate on failure states. Retaining offline demonstrations supplies nominal action supervision beyond those targeted corrections.

Online-only training (0:1) performs worse, with observed overfitting, drift, and stuck behavior. 1:2 is competitive with 1:1 but converges more slowly and slightly lower. The main experiments use fixed 1:1 sampling; this single-task ablation does not establish a globally optimal ratio.

Separately, LIFT w/ Offline DAgger uses a fixed force-correction buffer and underperforms across all tasks, scoring zero on book insertion despite force supervision. It removes repeated on-policy aggregation, whereas the 0:1 mixture removes offline demonstrations. These comparisons support complementary roles for offline coverage and continuously collected corrections.

How does LIFT handle tested shifts? Q5

The final LIFT checkpoint shows no clear degradation under the tested object, tablecloth, and lighting changes. These tests retain 10 rollouts per condition. The observation is limited to these shifts and does not establish broad task generalization or statistical equivalence.

LIFT performance under the original task conditions and tested object, tablecloth, and purple-light changes — **Performance under the tested shifts.** Original ten-rollout evaluation, reported separately from the n = 30 learning curves and ablations.

Photographs of the towel folding, book insertion, and Hanoi ring placement generalization setups — **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, with only the listed factor changing.

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

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}
}