acceptodds
Under review as a conference paper at ICLR 2027

The Surprising Effectiveness of Few-Step Adaptation for Robotic Manipulation

Abstract

Vision-language-action (VLA) models are pre-trained on large robot datasets and fine-tuned on hundreds of demonstrations per task, yet the natural way to teach a robot something new is to show it once. The prevailing answer for using that one demonstration is in-context learning, which retrains the policy to read demonstrations at inference. We ask whether something far simpler will do: keep the pre-trained VLA as it is and update it with a few gradient steps of its own training loss on the single demonstration. We call this Few-step One-Shot Imitation Learning (FOSIL). It works, and surprisingly well. One gradient step suffices when a familiar skill meets a new object or scene, ten for skills composed in a new order, and thirty for skills the policy could not perform before. We show this on a two-arm robot, with every result backed by 20 physical trials per task over 250 hours of evaluation, and on LIBERO-PRO, where the same demonstration given in context to the same backbone does not improve on the unadapted model. We then ask why so few steps are enough: robot pre-training decides the speed of adaptation, the useful update lives in a single direction per weight matrix, and a demonstration that moves differently from the policy can be refined with the policy’s own motion. A pretrained VLA, it turns out, is already a few gradient steps away from its next task. Physical robot rollouts and videos: https://fosilvla.github.io.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, robot manipulation, test-time adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8zraOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies one-demonstration adaptation of pretrained vision-language-action models. The models first undergo fine-tuning on pretext behaviors from the target setup. They then adapt to held-out tasks involving environmental changes, new compositions, or skills absent from the pretext set.

FOSIL updates all model parameters for 1-30 gradient steps using the original training objective. The paper studies the role of robot pretraining and the structure of the parameter updates. It also refines demonstration labels by combining policy-generated motion with demonstrated waypoint and gripper constraints.

Experiments cover LIBERO-PRO, compositional LIBERO tasks, and physical manipulation with π0.5\pi_{0.5} and MolmoAct2. FOSIL improves success over the pretext-trained models. It also achieves higher environmental suite averages and compositional success rates than the tested in-context baselines.

Critical strengths

The experiments show substantial improvements from a few updates on one demonstration. For π0.5\pi_{0.5}, average success on LIBERO-PRO task perturbations increases from 25% to 49% after one update. LIBERO-10 success increases from approximately 5% to 52% after ten updates. The results cover two pretrained VLAs and include physical robot evaluation. They establish that full-parameter fine-tuning is a strong baseline in this setting.

Critical weaknesses
  1. The adaptation gap is small in much of the evaluation. The models first receive 254 demonstrations on the same robot and setup, including 111 throwing demonstrations. Most test tasks modify or combine these trained behaviors. Only scooping and sweeping test skills absent from the pretext data. The results show that standard fine-tuning works well after this preparation, but provide limited evidence for learning substantially different behaviors from one demonstration.

  2. A low-rank update preserving performance does not show that the pretrained model already possessed the required skill. The refinement results also do not show whether the policy-generated actions are needed. The method smooths the actions and removes frames near grasps. The authors should check whether simply removing those frames or smoothing the demonstrated actions gives similar improvements.

Questions for authors
  • How much does adaptation depend on closely related pretext behaviors? For example, does adaptation to throwing new objects still work if throwing is excluded from pretext training, while the other behaviors are retained?
  • Does refinement still help when the baseline removes the same near-grasp frames but trains on the recorded actions? How does it compare with simply smoothing the demonstrated actions?
3 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We thank the reviewer for the careful reading, for emphasizing the few-step gains across two VLAs and on the physical robot, and for asking for the throwing hold-out and the refinement controls. Pointers below are to the revised PDF (blue).

"limited evidence for learning substantially different behaviors from one demonstration" … "How much does adaptation depend on closely related pretext behaviors?" … "does adaptation to throwing new objects still work if throwing is excluded from pretext training, while the other behaviors are retained?" (W1 / Q1)

With throwing removed from pretext, one demonstration still teaches the throw, but it takes about 30 steps instead of 1, and the soap lands in the strainer in 25% of trials.

Real-world throwing hold-out (App. G.5, Fig. 16). We retrained pretext without throwing and kept the other behaviors.

Pretext throwingMetric1 step10 steps30 steps64 steps128 steps
Included (Fig. 2)Task success (%)62————
Held out [new]Soap reaches target table (%)No throwNo throw557592
Held out [new]Soap lands in strainer (%)0002523
  • The policy learns the motion: it does not throw at 1 or 10 steps, and the soap reaches the target table in 55% of trials at 30 steps and 92% at 128.
  • Placement does not follow: 0% in the strainer at 30 steps, 25% at 64, and 23% at 128, so increasing the budget from 64 to 128 steps did not improve observed placement success.
  • Scooping and sweeping, whose motion is also absent from pretext, reach 60% and 30% at 30 steps (Fig. 2).
  • The 1-step, 62% throwing result in Fig. 2 is environmental adaptation of a pretext skill (111 of the 254 pretext demonstrations are throwing), not a new-skill claim. Fig. 2 is success across three objects; the hold-out rows are soap.

This dependence is discussed in Sec. 4 and App. G.5, where we highlight how robustness of adaptation for new-skills depends on task difficulty. Thus we hope that future models, with larger pretext behaviors will continue to improve few-step performances on even harder tasks like our more difficult novel skills learning included in the ablation.

New skills in simulation (App. G.3, Table 7, Fig. 13). We trained π0.5\pi_{0.5} on 78 LIBERO-90 behaviors, excluding all tasks involving turning stove knobs, opening microwave doors, and opening drawers. We then adapted from a single demonstration on the 12 held-out tasks, together with 20 environmental and 19 compositional tasks (shared grid 0, 1, 32, 64, 128 steps).

CategoryTasksMain paper budget (steps)Median steps to 50% success [new]
Environmental2012.3
Compositional19106.7
New skill123031.7
  • Eight of the twelve new-skill tasks reach 50% success; mean success across all twelve tasks at 32 steps is 46.8% (Table 7).
  • The shared grid has no points between 1 and 32, so the environmental and compositional medians are interpolated. App. G.3 does not find a clear separation between these two categories.

Compositional pretext hold-out (Table 9; App. G.5). Real-world success rates (%) per sub-task over 20 rollouts, with one demonstration and ten gradient steps from the π0.5\pi_{0.5} initialization:

MethodHandoverPlacementReach & pickZip
FOSIL (related pretext included)60508055
FOSIL (related pretext held out) [new]50352510

With the pretext behavior supplying the second sub-task held out, Zip falls from 55% to 10% and Reach & pick from 80% to 25%, while Handover falls from 60% to 50%.


"A low-rank update preserving performance does not show that the pretrained model already possessed the required skill." (W2)

The reviewer is right on this point: a low-rank update that preserves the adaptation gain is not evidence that the unadapted policy could already execute the skill. What the analysis shows is that robot pretraining makes the update compact:

  • LIBERO-Object: truncation at head-equivalent rank k=1k=1 preserves 96% of the adaptation gain for pretrained π0.5\pi_{0.5}, against 41% when the action expert is re-initialized (Fig. 5).
  • LIBERO-10, same number of steps and learning rate for both models: 87% of the gain against 52%.
  • Raw-gradient control: on three LIBERO-Object tasks, truncating the raw gradient at k=1k=1 before the optimizer step retains 92% of the gain (App. E).

This is consistent with Aghajanyan et al. (2021), who show that pretraining lowers the intrinsic dimension of downstream adaptation. Section 5.1 explicitly clarifies that the result does not show that the unadapted model already possesses the behavior (lines 408–410).


References: Aghajanyan et al. (2021), "Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning"

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

"Does refinement still help when the baseline removes the same near-grasp frames but trains on the recorded actions?" … "How does it compare with simply smoothing the demonstrated actions?" (W2 / Q2)

Refinement still helps. Removing the same near-grasp frames and training on the recorded actions gives 48.8% and 39.6%, against 76.2% and 75.2% for refinement. App. G.4 also reports a separate oracle pace-matching experiment and comparisons over larger adaptation budgets (Fig. 14).

The unadapted and recorded-demonstration rows are three-seed LIBERO-Object means (Table 8). The original single-seed recorded-demonstration results were approximately 52% / 36%.

Adapted on (LIBERO-Object, one step)Task perturbation (%)Object swap (%)
Unadapted9.510.3
Recorded demonstration47.537.3
Near-grasp frames removed [new]48.839.6
Self-refined demonstration76.275.2
  • Frame removal changes success by 1.3 and 2.3 points over the recorded demonstration, so it does not account for the gain.

  • Pace matching, separate run. App. G.4 reports resampling recorded trajectories at the policy's pace using oracle speed information from similar training demonstrations. This run reports pace-matched versus refined success of 56.0% versus 48.2% for task perturbation (p=0.23p=0.23), and 75.2% versus 72.2% for object swap. These results come from a separate run from the Table 8 comparison above. The oracle speed requirement also limits direct application to a new dataset.

  • Suboptimal demonstrations. To examine refinement beyond pace matching, we also evaluate adaptation on successful demonstrations whose execution differs from the training distribution. Appendix G.4 (Fig. 14) shows degradation with more steps on recorded demonstrations, while refinement preserves or improves success. Panel (a) covers five wrist-heavy LIBERO-Object tasks; panel (b) covers all ten LIBERO-Spatial task-perturbation tasks. Fig. 15 shows corresponding wrist-motion failures and refined rollouts.

  • Smoothing. The smoothness prior in App. C acts only on the correction that moves a policy sample onto the demonstrated waypoints; it does not filter the recorded actions. We did not low-pass or spline-smooth the recorded actions. The controls investigate frame removal and execution pace; they do not separately isolate the contribution of wrist orientation. The original schematic has been replaced by the illustration of task constraints and execution properties in revised Fig. 6(a).

Appendix G.4 discusses the contribution of pace and the differences that persist with more adaptation steps.

AuthorsOpenReview ↗

Official Comment by Authors

We are following up to ask whether our response addresses your concerns. To the best of our ability, we have addressed the questions through new results, additional experiments, and clearer interpretations:

  1. Dependence on related pretext behaviors (W1, Q1). We added the requested throwing hold-out, new-skill hold-outs in simulation, and compositional pretext ablations. These show that one demonstration can teach behaviors absent from pretext training, while related prior experience substantially improves adaptation speed and robustness. We explicitly distinguish acquiring a new motion from executing it reliably.
  2. Interpretation of low-rank updates (W2). We clarified that compact updates do not establish that the unadapted model already possessed the target skill. But rather supports the finding that pre-training enables effective adaptation through compact parameter changes, and connect its relation to previous work.
  3. Refinement controls and smoothing (W2, Q2). Additional experiments show that frame removal alone does not explain the refinement gains. Oracle pace matching can recover the one-step benefit in a separate comparison, while longer-budget experiments show benefits of refinement beyond pace on demonstrations with unusual wrist motion. We also clarified the role of the smoothness prior and replaced the schematic that could suggest filtering of recorded actions.

We posted our detailed responses on November 14 to allow time for further discussion. Is there any concern that remains unresolved? We are very open to running additional experiments during the remaining rebuttal period and providing any further detail.

Your questions led to direct tests of both the method’s capabilities and its limits. We believe this new evidence strengthens the paper’s conclusions and hope you will reconsider your assessment.

Reviewer_zBsXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether pretrained vision-language-action (VLA) models can adapt to new robotic manipulation tasks using a single demonstration and a few gradient updates. The authors propose FOSIL, which retains the original model architecture and training objective, and evaluate its effectiveness across environmental, compositional, and skill adaptation tasks in both simulation and real-world experiments. The paper further analyzes the role of robot pretraining, the low-rank structure of parameter updates, and demonstration quality in rapid adaptation, aiming to provide practical insights for future VLA adaptation research.

Critical strengths
  1. Demonstrating an effective and simple approach to one-shot adaptation. The paper proposes Few-step One-Shot Imitation Learning (FOSIL), which adapts pretrained VLA models using a single demonstration and a few gradient updates, without modifying the original architecture or training objective. Experiments across different VLA models, simulation benchmarks, and real-world robotic manipulation tasks demonstrate the practical effectiveness of this simple approach.
  2. Investigating the relationship between task novelty and adaptation speed. The paper categorizes adaptation tasks into environmental adaptation, compositional adaptation, and skill adaptation, demonstrating performance improvements using 1, 10, and 30 gradient steps, respectively. The authors further investigate the effects of pretrained initialization, parameter update structure, and demonstration quality, providing potentially useful insights into rapid VLA adaptation.
Critical weaknesses
  1. Insufficient clarification of the novel insights beyond prior work. One-shot adaptation of robotic policies and pretrained VLAs has already been explored in prior studies. Although this paper differs in its adaptation mechanism and experimental settings, it remains unclear what fundamentally new insights its empirical findings and analyses provide beyond demonstrating the effectiveness of standard gradient updates, and how these insights can guide future research.
  2. Insufficient evidence for a general relationship between task novelty and adaptation steps. The paper associates environmental, compositional, and skill adaptation with 1, 10, and 30 gradient steps, respectively. However, the criteria for selecting these budgets are not clearly explained. Figure 3 shows that some environmental adaptation tasks continue to improve substantially with additional gradient steps, while complete adaptation curves for skill adaptation are missing. Therefore, the current evidence supports differences in adaptation speed across tasks but does not sufficiently establish a generalizable relationship between task novelty and the required number of gradient steps.
Questions for authors

The paper builds on the empirical observation that pretrained VLAs can adapt to new tasks through a few gradient updates on a single demonstration. It further analyzes this phenomenon and derives several practical insights. For this type of empirical study, a key consideration is whether the findings provide new knowledge beyond prior work and whether the resulting insights can reliably guide future research. Q1. Novelty and significance of the central finding. Prior studies, including One-Shot Visual Imitation Learning via Meta-Learning (CoRL 2017), MoS-VLA (2025), and VLA Models Are More Generalizable Than You Think (CVPR 2026), have already demonstrated rapid one-shot adaptation of robotic policies or pretrained VLAs from different perspectives. Although these studies differ in their adaptation mechanisms and task settings, what new insights does this paper establish beyond demonstrating the effectiveness of standard gradient updates? In particular, how do the findings on behavioral adaptation and skill composition provide new guidance for future research? Q2. Reliability of the relationship between task novelty and adaptation steps. The paper associates environmental, compositional, and skill adaptation with 1, 10, and 30 gradient steps, respectively, and provides partial adaptation curves in Figure 3. However, the selection criteria and generalizability of these budgets remain unclear. Could the authors explain how these numbers were determined and provide additional adaptation curves, particularly for skill adaptation, to demonstrate a consistent relationship between task novelty and adaptation speed rather than task-dependent choices of training budgets? Can this relationship guide the selection of adaptation steps for unseen tasks?

Additional Comments Q3. Missing supplementary material. The manuscript repeatedly refers to Appendices A–F and Figure 7, but these materials are not included in the provided PDF. They contain important details about the experimental setup, training procedures, and supporting analyses. Please clarify whether they were submitted separately and ensure they are accessible for review and reproducibility. Q4. Code availability. The Reproducibility Statement indicates that the code and rollout videos are available at https://fosilvla.github.io. However, the project page does not appear to provide an accessible code repository or download link. Please clarify how the code can be accessed to reproduce the reported results.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We thank the reviewer for describing FOSIL as "an effective and simple approach to one-shot adaptation," and for asking whether the 1, 10, and 30-step split can actually guide an unseen task. That was a useful request: we added skill-adaptation curves and a throwing hold-out, both below.

"it remains unclear what fundamentally new insights its empirical findings and analyses provide beyond demonstrating the effectiveness of standard gradient updates, and how these insights can guide future research" (W1 / Q1)

Finn et al., MoS-VLA, and Li et al. each show that one-shot adaptation is possible with a mechanism built for it: a meta-trained initialization, a gradient-free update, and visual-token adapters for viewpoint. We ask a different question: what can a standard pretrained VLA already do with plain gradient steps, and why? Finding 1 in Sec. 4 and Insights 1–3 in Sec. 5 are that answer. Each comes with guidance for future work.

Finding 1. Robot VLA pretraining, not meta-training, is what makes the first gradient step large. Across four initializations trained on the same pretext behaviors, robot-pretrained models gain the most from step 1. On LIBERO-Object, one step takes π0.5\pi_{0.5} from 9.6% to 52.0%; the same model with its action expert re-initialized goes from 5.8% to 8.2% (Fig. 4). Finn et al. meta-train an initialization to get that first step; MoS-VLA avoids gradients. Guidance: if the initialization is a robot-pretrained VLA, start from few steps of the original loss. (Finding 1, Sec. 4, lines 381–384.)

Insight 1. The adaptation update is compact, and its effect is concentrated in the early VLM layers. At head-equivalent rank k=1k=1 (defined in App. E), truncation retains 0.96 of the success gain on LIBERO-Object and 0.87 on LIBERO-10; with the action expert re-initialized it retains 0.41 and 0.52 (Fig. 5, App. E). Rank-16 LoRA, whose first step lives in a random subspace, gains little on the environmental averages (Table 1). Guidance: for a change of object, goal, or skill, a visual adapter or a random LoRA subspace is not a substitute for the directions the pretrained model already has. (Insight 1, Sec. 5.1, lines 415–419.)

We also provide insights on how the locality of updates (Appendix: E2) differ between pre-trained and non-pretrained models in-addition to the geometry that we highlight in main paper. We hope that, this analysis of geometry and locality of updates serve as useful observation for designing as well as selection of models that adapt fast and robustly.

Li et al. adapt visual tokens for viewpoint, over many adapter steps on one demonstration. We implemented FTM and FLA on the environmental tasks of Table 1 with π0.5\pi_{0.5}, the same demonstrations, and the 1-step budget we use, over 3 seeds. Both stay at the unadapted policy.

MethodTask perturbation success (%)Object swap success (%)
Unadapted π0.5\pi_{0.5} (3-seed)25.620.0
[new] FTM, 1 step26.520.4
[new] FLA, 1 step26.721.7
FOSIL, 1 step (3-seed)44.033.6

That is consistent with Fig. 10, where the VLM carries 63% of the one-step effect and SigLIP 25%. The original submission reported a single-seed task-perturbation average of approximately 49%; revised Table 1 reports a three-seed mean of 44.0%. The related-work distinction is in Sec. 2, and the direct comparisons are in Sec. 4 and App. G.2 (blue).

Insight 2. A demonstration can specify the right task and still limit few-step success through how it is executed (pace, wrist). Adapting on the refined demo rather than the recorded one lifts LIBERO-Object one-step success from 47.5% to 76.2% under task perturbation and from 37.3% to 75.2% under object swap (Fig. 6b; Table 8). We also do detailed ablations behind these gains in App. G.4.

Guidance: distributional similarity between train time and test-time demos guide robustness of adaptation, and self-refinement is preliminary exploration into sample from policy to bridge the gap at test-time.


References:

[1] Finn et al., "One-Shot Visual Imitation Learning via Meta-Learning", CoRL 2017

[2] Zhao et al., "MoS-VLA: A Vision-Language-Action Model with One-Shot Skill Adaptation", arxiv 2025

[3] Li et al., "VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling", CVPR 2026

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Insight 3. On one backbone, under environmental adaptation, few-step adaptation improves success while the tested in-context methods provide little improvement over the unadapted policy. The following values are from revised Table 1; the unadapted and FOSIL rows are three-seed means, while ICL and RiCL retain their original evaluation.

MethodTask perturbation success (%)Object swap success (%)
Unadapted π0.5\pi_{0.5} (Table 1)25.620.0
ICL [π0.5\pi_{0.5}] (Table 1)20.718.3
RiCL [π0.5\pi_{0.5}] (Table 1)25.019.3
FOSIL [π0.5\pi_{0.5}] (Table 1)44.033.6

ICL and RiCL stay at the unadapted model under task perturbation; FOSIL does not. ICL needs training relations that make the demonstration necessary; more ordinary robot data does not supply them (Fig. 7). Guidance: do not assume a context pathway on a pretrained VLA will read the demonstration. Evaluate the two routes on one backbone. (Insight 3, Sec. 5.3, lines 505–508.)


"how do the findings on behavioral adaptation and skill composition provide new guidance for future research?" (W1 / Q1)

The usable check is whether the skill already appears in the pretext behaviors.

  • If the skill is already in pretext, a few gradient steps retarget object, layout, or order (environmental and compositional in the paper).
  • If it is not, budget tens of steps, and expect the motion to appear before spatial robustness (skill adaptation).

We suggest that few-shot VLA studies report that distance alongside the step budget.

We tested this by holding the skill out of pretext and adapting from one demonstration. On the robot we keep the other pretext families and remove throwing (App. G.5, Fig. 16). With throwing in pretext, FOSIL (π0.5\pi_{0.5}) in Fig. 2 reaches 62% at one step across three objects. The table below uses a soap bar, with throwing held out of pretext.

1 step10 steps30 steps64 steps
[new] Throwing held out, reaches table (%)no throwno throw5575
[new] Throwing held out, lands in strainer (%)00025

The throwing motion appears around 30 steps (55% reach the table), but landing in the strainer is still only 25% at 64 steps. So 30 steps is where a new motion appears; how reliable it becomes depends on the skill (scooping reaches 60% at 30 steps in Fig. 2). Videos are on the project webpage.

The same split shows up in simulation: 12 new-skill hold-outs from one π0.5\pi_{0.5} trained on 78 LIBERO-90 behaviors (stove knobs, microwave doors, and drawers held out) have a median of 31.7 steps to 50% success (8/12 reach 50%). Environmental and compositional tasks in the comparison adapt faster. The new-skill model was retrained with the skill categories held out, as described in App. G.3. Full curves: Fig. 13, Table 7, App. G.3.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

"the selection criteria and generalizability of these budgets remain unclear" … "provide additional adaptation curves, particularly for skill adaptation" … "Can this relationship guide the selection of adaptation steps for unseen tasks?" (W2 / Q2)

The initialization analysis (Fig. 4 in the revised PDF) shows continued improvement beyond one step. We have added new-skill adaptation curves in App. G.3. That analysis supports slower new-skill adaptation, while the appropriate budget also depends on the task and prior experience.

This is a great suggestion. We implemented it. Curves to 128 steps on 51 tasks (20 environmental, 19 compositional, 12 new-skill) are Fig. 13, Table 7, App. G.3 of the revised PDF; the throwing hold-out is Fig. 16, App. G.5. Thanks again for the recommendation to strengthen the adaptation-speed analysis.

We evaluate on a shared grid of 0, 1, 32, 64, and 128 steps and interpolate the step count at 50% success (App. G.3).

Adaptation categoryNumber of tasks[new] Median gradient steps to 50% success ↓\downarrow
[new] Environmental202.3
[new] Compositional196.7
[new] New skill (held-out categories)1231.7 (8/12 tasks reach 50%)

However, we would like to maintain that robustness varies inside each categories and environment: scooping, a skill absent from pretext, is much more reliable than throwing with throwing held out (Figure 16). And also adaptation in real-world with limited diversity might need more steps to overcome visual shortcuts of pretext behaviors like our real-world experiments that are not captured faithfully in our simulation evaluations.

How 1, 10, and 30 were chosen.

  • They were chosen heuristically from preliminary loss convergence rather than tuned for maximum success.
  • The same three values are used on every suite, with no per-task tuning.
  • Step count matters more than learning rate: on LIBERO-90 held-out, moving from one to ten steps changes success from 21.1 to 66.4, a larger spread than the three learning-rate candidates (58.9–66.4).

For an unseen task, our practical heuristic is to stop when demonstration loss plateaus, following the approach used to choose the budgets in our preliminary runs. This is a starting point rather than a guarantee of maximum success or robustness. The reported budgets indicate what current models can do, rather than the optimal choice for every unseen task.


"The manuscript repeatedly refers to Appendices A–F and Figure 7, but these materials are not included in the provided PDF." (Q3)

The appendices were submitted as a separate supplementary PDF. We have now incorporated Appendices A–F and the new Appendix G into the revised main PDF. The pretext-behavior illustration is Fig. 8 in this version, so the referenced material is accessible in one document.


"the project page does not appear to provide an accessible code repository or download link" (Q4)

Thank you for raising this. We have posted the code, checkpoints, and datasets at https://github.com/fosilvla/fosilvla-code. The same URL is now in the Reproducibility Statement.

AuthorsOpenReview ↗

Official Comment by Authors

We are following up to ask whether our response addresses your concerns. To the best of our ability, we have addressed all questions with clarifications, new results, and additional experiments:

  1. New insights beyond prior work (W1, Q1). We clarified how pretraining shapes the geometry and locality of adaptation updates, how demonstration–policy compatibility affects robustness, and why effective gradient-based adaptation does not imply effective in-context adaptation. New baseline comparisons and refinement controls strengthen these findings and their implications for model selection, demonstration design, and evaluation.
  2. Task novelty and adaptation speed (W2, Q2). We added adaptation curves covering environmental, compositional, and new-skill tasks, together with a real-world throwing hold-out. These show that acquiring skills absent from pretext training generally requires more updates, and that acquiring the motion can precede robust execution. We also clarified that the original budgets were heuristic choices, not success-tuned optima or universal thresholds.
  3. Supplementary material (Q3). The appendices were submitted separately. We have now incorporated them into the revised main PDF so that all experimental details and supporting analyses are accessible in one document.
  4. Code availability (Q4). We released the code, checkpoints, and datasets, and added the repository link to the Reproducibility Statement.

We posted our detailed responses on November 14 to allow time for further discussion. Is there any concern that remains unresolved? We are very open to running additional experiments during the remaining rebuttal period and providing any further detail.

Your questions helped us sharpen both the evidence and the scope of our claims. We believe the new results materially strengthen the paper and hope you will reconsider your assessment.

Reviewer_at36OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the behavior of few-shot adaptation of VLAs (from even one demonstration) just through gradient descent after supervised tuning. Somewhat surprisingly, they find that this can adapt to new manipulation tasks with one demonstration and 30 gradient steps. While there is past work that studied this kind of implicit meta-learning and rapid adaptation with few gradient steps, the authors show that this can work even with an extremely low step budget using a large VLA model (π0.5\pi_{0.5}). Beyond the interesting empirical finding, the authors analyze how initialization and low-rank weight updates enable this rapid demonstration adaptation.

Critical strengths
  • The main strength of the paper is the somewhat surprising result that you can get such good performance with only a few steps. The authors present clear real-world experiments showing that this works even with fewer than 30 steps, which is better than prior work.
  • Other strengths include the diverse multi-arm real-world and simulated experiments.
  • Ablations show that natural approaches like in-context learning (ICRT [1], RiCL [2], and BPP [3]), as well as rank-16 LoRA [4], are not as effective here under the standard pretext-training regime. The same-backbone ICL and LoRA baselines do improve over the unadapted model on compositional tasks, and context-informative training data substantially improves the in-context methods.

References

[1] Fu et al. In-Context Imitation Learning via Next-Token Prediction. ICRA, 2025.

[2] Sridhar et al. RICL: Adding In-Context Adaptability to Pre-Trained Vision-Language-Action Models. CoRL, 2025.

[3] Patel et al. What Enables In-Context Behavior Prompting for Manipulation?. arXiv, 2026.

[4] Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR, 2022.

Critical weaknesses
  • One key concern is novelty relative to prior few-shot adaptation works based on gradient descent. Can you clarify how the approach relates to these past methods: Finn et al. [1] and Li et al. [2]? Please compare. It would be good to directly evaluate against Feature Token Modulation (FTM) and Feature Linear Adaptation (FLA) from Li et al. [2].
  • A major concern is uncertainty about the results across independent training runs and demonstration choices. Clarify whether results were averaged across different training runs and whether demonstrations varied across runs. There are fixed evaluation seeds in Appendix B and confidence bands in Fig. 4, but the paper would benefit from more rigorous analysis of uncertainty across the board.

If these concerns are addressed, I'm happy to raise my score.

References

[1] Finn et al. One-Shot Visual Imitation Learning via Meta-Learning. CoRL, 2017.

[2] Li et al. VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling. CVPR, 2026.

Questions for authors
  • Is the code available? The paper states that the implementation is made available, but there doesn't appear to be a link in the text.
  • Could you clarify the relationship to Li et al. [1], comment on novelty, and directly compare against their approach?
  • Could you clarify why the Figure 5(b) perturbation results seem inconsistent with the numbers reported in Appendix C? The experiment descriptions look similar, but the 8-step adaptation results are reported differently.

References

[1] Li et al. VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling. CVPR, 2026.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We thank the reviewer for highlighting the few-step real-world result and the ICL / LoRA ablations. Those were the parts we most hoped a reader would take away.

Thank you for offering to raise the score after we address the listed concerns, and for so clearly spelling out what would address them. We’ve included new results to address them objectively:

  • Relation to Finn et al. and Li et al., including FTM / FLA on the Table 1 π0.5\pi_{0.5} settings and on their viewpoint suite. At one step, FTM, FLA, and LoRA remain close to the unadapted baseline, while FOSIL provides substantially larger gains. That strengthens Finding 1 / Fig. 4, and is consistent with our module-level analysis (Fig. 10) of where effective one-step updates occur.
  • Three seeds for Table 1, plus three demonstrations. Object, Goal, LIBERO-10, and LIBERO-90 stay large across seeds; demo spread is a few points. That strengthens the Table 1 few-step result. Spatial at one step does not (Insight 2). Pretext-training seeds will be added in a revision.

We are happy to address any follow-up quickly and with any additional results that may further substantiate our response.


"how the approach relates to these past methods: Finn et al. [1] and Li et al. [2]"

FOSIL does not meta-train for few-shot adaptation. Finn et al. learn an initialization for one-shot imitation. We start from a standard pretrained VLA, train on pretext behaviors with the evaluation tasks held out, and take a few steps of the original loss on one demonstration (Sec. 3.1). Finding 1 / Fig. 4: VLA pretraining makes the first of those steps large; re-initializing the π0.5\pi_{0.5} action expert removes most of that first-step gain. Li et al. adapt the vision encoder for viewpoint changes; our direct comparisons are below and in App. G.2 (Tables 5–6). This distinction is discussed in Sec. 2 (blue).


"directly evaluate against Feature Token Modulation (FTM) and Feature Linear Adaptation (FLA)"

We implemented the suggestion made in the review. FTM and FLA on the Table 1 π0.5\pi_{0.5} settings, same demos, batch, and 1-step budget (3 seeds). Per-suite cells are in App. G.2; Spatial is unchanged at one step.

MethodTask pert. mean (%)Object swap mean (%)
Unadapted (Table 1)25.620.0
[new] FTM26.520.4
[new] FLA26.721.7
LoRA r16 (Table 1)26.620.9
FOSIL (Table 1)44.033.6

At one gradient step, FTM, FLA, and LoRA stay at the unadapted policy. FOSIL does not. That is consistent with Fig. 10 / App. E.1: the useful one-step update is concentrated in the VLM, not a visual-token shift. Table 1 and App. G.2 (blue).

On Li et al.'s own viewpoint suite, 10 FOSIL steps match 300 FTM steps. The first column averages task perturbation and object swap at the step budget specified in each row (Table 1 for 0/1 steps; Table 5 for 10 steps). The viewpoint results are from Table 6 and were not in the original submission.

MethodEnvironmental task/swap mean (%)Li et al. viewpoint suite, 10 steps (%)Li et al. viewpoint suite, 300 steps (%)
Unadapted, 0 steps (Table 1)22.8
FOSIL, 1 step (Table 1)38.8
[new] FTM, 10 steps25.961.474.4
[new] FLA, 10 steps45.773.881.0
[new] FOSIL, 10 steps53.174.5not run

These results highlight that simple fine-tuning on strong pre-trained VLAs are enough for robust adaptation and serve as a strong baseline for specialized adapters research that the community will benefit from and help in progress on not just test-time procedures but how to improve pre-training objectives/architectures themselves for more strong adaptation.

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

"uncertainty about the results across independent training runs and demonstration choices"

We completely agree. We report three seeds for the π0.5\pi_{0.5} Table 1 numbers and three demonstrations on Object task-pert, Goal object-swap, and LIBERO-10. Each seed varies the adaptation randomness and evaluation episodes; the pretrained checkpoint and demonstration remain fixed across seeds (App. G.1). Pretext-training seeds will be added in a revision. Environmental suites: 1 step. Compositional: 10 steps.

The following entries are adaptation gains in percentage points; three-seed entries report mean ± SD (Table 4). The single-seed columns refer to the original submission.

SuiteStepsOriginal task (1 seed)Original swap (1 seed)[new] 3-seed task[new] 3-seed swap
LIBERO-Object1+42+42+26+2638.0±3.338.0 \pm 3.327.0±2.427.0 \pm 2.4
LIBERO-Spatial1+7+7+5+51.2±3.41.2 \pm 3.41.8±4.01.8 \pm 4.0
LIBERO-Goal1+22+22+12+1216.1±2.916.1 \pm 2.911.8±4.211.8 \pm 4.2
LIBERO-1010+47+4750.3±3.550.3 \pm 3.5
LIBERO-9010+48+4848.9±2.948.9 \pm 2.9

Object, Goal, LIBERO-10, and LIBERO-90 stay large across seeds. Spatial at one step does not (the submission's 51 →\to 58 was a single seed). App. G.1 (blue).

Original demonstration: Object 47.5 and Goal swap 25.9 are the revised Table 1 three-seed means; LIBERO-10 is the original single-seed result, rounded to 52%. The SD column below is calculated only across demonstrations A/B/C.

SettingOriginal demo (%)[new] Demos A / B / C (%)Spread (SD, pp) ↓\downarrow
Object task perturbation47.554.8 / 57.6 / 63.24.3
Goal object swap25.925.1 / 27.1 / 27.71.4
LIBERO-105251.6 / 50.4 / 57.43.7

Across the three tested demonstrations, SD ranges from 1.4 to 4.3 percentage points. This is separate from the adaptation-and-evaluation variability measured across seeds.

We also examine demonstration variability in real-world tasks using suboptimal demonstrations. Fig. 12 shows a handover demonstration in which the object is dropped; FOSIL reproduces this behavior. Our current refinement procedure cannot correct such errors, which remain an opportunity for future work.

"Is the code available?"

Yes. https://github.com/fosilvla/fosilvla-code (FTM / FLA included). Now also in the Reproducibility Statement.

"clarify the relationship to Li et al." … "directly compare against their approach"

Please see the two FTM / FLA tables above.

"Figure 5(b) perturbation results seem inconsistent with the numbers reported in Appendix C"

Thank you for catching this. The figure reported the best scores across adaptation budgets for recorded and refined demonstrations, whereas Appendix C compared results at matched budgets. We have now included the full adaptation curves for the suboptimal demonstrations for LIBERO-Spatial in Appendix G.4 (Fig. 14(b)), and the revised refinement panel, Fig. 6(b), reports the one-step LIBERO-Object comparison from Table 8.

AuthorsOpenReview ↗

Official Comment by Authors

We are following up to ask whether our response addresses your concerns. To the best of our ability, we have addressed each question with direct evidence, including new results and additional experiments:

  1. Novelty and comparisons with FTM/FLA (W1, Q2). We added direct comparisons under matched adaptation budgets and on their viewpoint benchmark. These strengthen our central finding: standard pre-trained VLAs support rapid behavioral adaptation through full fine-tuning, without meta-training or specialized adapters.
  2. Reliability across runs and demonstrations (W2). Additional experiments show that the main gains persist across adaptation-and-evaluation seeds and demonstration choices. They also clarify the limited improvement on LIBERO-Spatial and the effect of demonstration imperfections.
  3. Code availability (Q1). We released the implementation, including our FTM/FLA baselines, and added the repository link to the Reproducibility Statement.
  4. Figure consistency (Q3). We corrected the refinement comparison to use matched adaptation budgets and added full adaptation curves to make the comparison transparent and ablations to identify the components responsible for improvement under our refinement.

We posted our response with time for further discussion and would welcome your assessment of whether any concern remains unresolved. We are also open to running additional experiments during the remaining rebuttal period and providing any further detail.

Your review helped sharpen both the contribution and its limits. We believe the new evidence materially strengthens the paper and hope you will reconsider your assessment.

Reviewer_at36OpenReview ↗

Response to author rebuttal

Thank you for the thoughtful response. The comparisons and code release address several of my concerns. The additional adaptation/evaluation seeds and demonstration checks also help. I am keeping my rating at Weak Reject.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Authors

As the discussion period comes to an end, we thank reviewers at36, zBsX, and 8zra for their feedback. We posted detailed responses on November 14 and follow-ups on November 17, with substantial new experiments addressing their concerns. Across the reviews, several core strengths of the work were recognized:

  • Simple and effective adaptation (at36, zBsX, 8zra): “somewhat surprising result,” “an effective and simple approach to one-shot adaptation,” and “full-parameter fine-tuning is a strong baseline in this setting,” respectively.
  • Strong experimental evaluation (at36, zBsX, 8zra): All three reviewers recognize the evaluation across multiple models, simulation, and physical robots. at36 specifically highlights the “diverse multi-arm real-world and simulated experiments.”
  • Useful analysis and insights (at36, zBsX): The reviews highlight our investigation of pretraining, model updates, and demonstration quality, with zBsX recognizing “potentially useful insights into rapid VLA adaptation.”

In response to reviewer suggestions, we have made substantial efforts to objectively address the concerns through new experiments and clearer explanations. These additions strengthen our submission both empirically and conceptually:

  • Direct baseline comparisons (at36, zBsX): Added FTM/FLA evaluations on our tasks and their viewpoint benchmark, demonstrating stronger few-step behavioral adaptation and competitive viewpoint adaptation using standard fine-tuning.
  • Seed and demonstration robustness (at36): Repeated experiments across seeds and demonstration choices, confirming substantial gains on LIBERO-Object, LIBERO-Goal, and the compositional suites.
  • Additional new-skill evaluations (zBsX, 8zra): Extended adaptation curves across 51 tasks and added real-world throwing and compositional experiments with related training behaviors removed, directly testing learning beyond previously practiced skills.
  • Further analysis of pre-training (zBsX, 8zra): Added controlled comparisons that substantiate how robot pre-training enables effective adaptation through small, focused changes to the model.
  • Extensive refinement ablations (8zra): Added controls for frame removal and execution speed, together with experiments on demonstrations with unusual wrist motion, demonstrating refinement’s benefits across different forms of suboptimal execution.

In-addition, we also clarified the adaptation budgets and robustness based on task difficulty, and released code, checkpoints, and datasets.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.