Never Too Late for Force: Accelerating VLA Post-Training with Reactive Force Injection

LIFT is a force-aware post-training framework for pretrained VLA policies that adds contact reactivity while aiming to preserve the pretrained manipulation prior.

Source: [12], [28]

The principal vision-only comparison is π0.5 with online DAgger, which uses the same online DAgger loop without force input.

Source: [5]

Other comparisons include LIFT with single-frame force instead of reactive force memory, LIFT with a fixed offline correction buffer, π0.5 trained with offline handheld data only, a lightweight residual policy, and full LIFT.

Source: [4], [16], [21], [26], [29]

The offline dataset is vision-only task-alignment data collected with a handheld device; the online dataset contains force-enabled corrections collected on states visited by the deployed policy.

Source: [7], [24]

The real-robot rollouts were executed on a Flexiv Rizon 4S with a 6D end-effector force sensor; the paper states that its experiments used a single-arm platform.

Source: [17], [33]

For towel folding, a reported failure is that visual depth ambiguity can prevent the policy from determining whether the gripper has contacted or grasped the thin cloth.

Source: [3]

For book insertion, reported failures include pushing after the book has bottomed out or releasing before it is seated, which can cause the book to flip or fail to settle.

Source: [18]

For Hanoi ring placement, misalignment can jam the ring; the paper also reports that single-frame-force behavior can become unstable after a collision, including oscillation, loss of localization, or movement away from the placement region.

Source: [1], [27]

On book insertion, LIFT reached a peak score of 0.6 at 4.6K samples, compared with 0.4 at 5.6K samples for π0.5 with online DAgger.

Source: [18]

On Hanoi ring placement, LIFT reached and maintained a score of 0.6 at the final checkpoint (2.0K samples); π0.5 with online DAgger peaked at 0.3 and ended at 0.2.

Source: [27]

The paper reports that LIFT showed no substantial degradation in most tested shifted settings involving objects, tablecloths, and lighting; this reported association does not establish that the method preserves generalization in untested conditions.

Source: [23]

On book insertion, LIFT reached a peak score of 0.6 at 4.6K samples, while the single-frame-force variant peaked at 0.4; on towel folding, the paper reports comparable performance between the single-frame-force and reactive variants.

Source: [6], [15]

The paper reports that LIFT with offline DAgger underperformed on all three tasks and fell to zero on book insertion, while online DAgger adds corrections on learner failure states.

Source: [8], [19]

The tested lightweight residual policy achieved substantially lower scores than LIFT; the authors attribute this in part to large required corrections from the weak base and sparse residual targets that make intervention timing difficult to learn.

Source: [2], [20]

The experiments were conducted only on a single-arm platform; the authors identify bimanual tasks on dual-arm platforms and other force sensors as future directions.

Source: [33]

The authors note that post-training the full VLA carries substantial computational cost.

Source: [2]

The reactive action stream decodes actions causally within a chunk, and the force pathway uses recent 6D end-effector force encoded as causal force memory with a latency-aligned causal mask.

Source: [31], [32]

LIFT caches the vision-language prefix and reuses it while reevaluating the action experts with updated force history within an action chunk.

Source: [25]

At initialization, copied action-expert weights and zero-initialized cross-attention output make the reactive force residual ineffective, so the augmented model starts with the same action output as the original model.

Source: [13], [22]

Towel folding and book insertion use stage-based scores, while Hanoi ring placement uses binary success.

Source: [17]

Each checkpoint was evaluated with three groups of ten autonomous rollouts, and 95% confidence intervals were computed; the reported online sample counts represent human-intervention frames, the only online data used for training.

Source: [17]

The paper asks how to inject force so a VLA can react quickly to contact while retaining force memory, how to preserve the pretrained VLA prior during adaptation, and how to post-train effectively when policy-dependent force distributions shift beyond offline coverage.

Source: [30]

The evaluation covered three real-robot manipulation tasks: towel folding, book insertion, and Hanoi ring placement.

Source: [17]

The reported shifted-condition evaluation varied objects, tablecloths, and lighting; the paper reports results for most tested settings but does not claim substantial degradation was absent in every setting.

Source: [23]