acceptodds
Under review as a conference paper at ICLR 2027

Path Compatibility: Persistent Anchors for Continual VLA Fine-Tuning

Abstract

Pretrained vision–language–action (VLA) policies provide a foundation for robots to acquire new skills throughout deployment. Continual fine-tuning, however, changes the perception-to-action path and can disrupt previously learned behavior. Distillation from a previous-task teacher uses a reference that may already have drifted from the response learned when an old task was acquired. We introduce Path Compatibility (PathComp), which preserves old responses together with the interface inputs that produced them. After learning each task, PathComp caches modal tokens together with their fused latents and action distributions. During later tasks, the current fusion module and action head reproduce these acquisition-time responses from the cached tokens, while a separate modal anchor constrains the encoders on stored observations. This construction provides a fixed downstream reference while accounting for upstream drift, without a rolling teacher. We evaluate PathComp with a CLIP–LoRA–GMM policy on LIBEROObject, LIBERO-Goal and LIBERO-Spatial, spanning changes in manipulated objects, task goals and spatial relations. Across these suites, it reduces final held-out action MSE by 21.5–22.6% relative to the strongest evaluated prior-work baseline under the shared policy and offline protocol in each suite. On LIBERO-Object, a shared-design target-time control and component ablations support the use of acquisition-time interface relations for retaining learned responses as the policy adapts. Code and checkpoints will be made public after acceptance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
continual learning, vision-language-action models, lifelong imitation learning, catastrophic forgetting, knowledge distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_s3ArOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PathComp addresses forgetting in continual VLA fine-tuning by preserving acquisition-time input–response relations. It caches modal tokens, fused latents, and action distributions, then constrains the current policy to reproduce these responses while separately regularizing encoder drift on stored observations. Offline evaluations on three LIBERO suites show reduced action prediction error compared with the evaluated baselines.

Critical strengths
  • The paper addresses policy drift in continual VLA fine-tuning.
  • Persistent interface anchors provide an intuitive approach to preserving learned responses.
  • Experiments across three LIBERO suites and component ablations support the method’s effectiveness in offline action prediction.
Critical weaknesses
  • The main experiments evaluate MSE and NLL on demonstration observations without closed-loop execution. But lower offline error does not necessarily imply higher task success.
  • Evaluation on a longer task sequence, such as LIBERO-90, would provide stronger evidence of continual-learning capability.
  • As tasks accumulate, the shared modules must accommodate representations and targets cached at different learning stages. Potential conflicts among these historical anchors raise concerns about long-term scalability.
  • The lack of real-robot experiments limits evidence of the method’s effectiveness in practical deployment.
Questions for authors

See weakness above.

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

This paper studies how robot policies can learn new tasks without forgetting old ones. After learning each task, PathComp saves intermediate features and action distributions for a small set of examples. During later training, it uses the saved modal tokens as inputs and matches the saved fused features and action distributions. The encoders are also trained to produce similar features for the stored observations as before. With a shared CLIP–LoRA–GMM policy, PathComp achieves lower offline action MSE than the tested baselines on three LIBERO suites. The paper does not establish whether these gains translate into better closed-loop task performance.

Critical strengths
  1. S1: PathComp gives the model a stable reference for retaining old tasks. A teacher from the previous task may already have forgotten some older tasks. PathComp instead keeps the targets saved when each task was learned. It also handles encoder drift separately from changes in the fusion module and action head, giving each loss a clear role.

  2. S2: The offline results support the method. On LIBERO-Object, PathComp has lower final action MSE than M2Distill-R and DER++-style replay in all three seeds. Using previous-task targets instead of fixed targets raises MSE from 0.341 to 0.400. The average action error at the end of each task is also similar to the baselines, so the method retains old tasks without an obvious loss in new-task prediction accuracy.

Critical weaknesses
  1. W1: The paper does not show whether PathComp helps robots retain old skills during actual execution. The main policy is evaluated on held-out demonstrations, without closed-loop rollouts. The OpenVLA experiment includes rollouts, but the training and evaluation task lists do not match, and there is only one adaptation stage. It also shows no clear improvement over sequential fine-tuning. The authors acknowledge these issues, but lower offline error alone is not enough to show better skill retention during execution.

  2. W2: The results support keeping fixed targets, but the benefit of replaying cached tokens is still unclear. The paper needs a direct comparison that keeps the same targets and losses but recomputes tokens from stored observations with the current encoders. The replay baselines use different losses and training settings, so the gains cannot be attributed to cached tokens alone. PathComp also received tuning on Object, while the baseline weights were taken from prior work or left at their default values. A more controlled comparison is needed to show whether caching the tokens helps.

Questions for authors
  1. Q1: How does the main policy perform in closed-loop rollouts on old and new tasks? For OpenVLA, do you have results after fixing the task-list mismatch and training through more than one adaptation stage?

  2. Q2: Does using cached tokens work better than recomputing tokens from stored observations with the current encoders? Could you compare these two options using the same fixed targets, losses, and other settings?

  3. Q3: Does PathComp still outperform the baselines after tuning their loss weights? Does confidence weighting still help if every anchor in the uniform-weight baseline is given a weight of 0.525, matching the average weight in PathComp?

  4. Q4: M2Distill-R has better held-out NLL but worse action MSE. Does PathComp still have lower MSE if actions are decoded using the mixture mean or the mean of the component with the highest mixture weight?

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

The paper studies continual fine-tuning of pretrained vision-language-action (VLA) policies. Continual fine-tuning changes the perception-to-action path: encoders, temporal fusion, and action head drift, so a previous-task teacher no longer reflects the response learned when an old task was acquired, and distillation from the current predecessor cannot recover that drifted behavior.

Critical strengths
  1. Precise, under-addressed problem with a traceable anchor decomposition. PathComp targets the fact that the teacher reference itself drifts during continual VLA adaptation, an issue most distillation-based lifelong imitation methods overlook. The decomposition into a core policy anchor, an auxiliary latent anchor, and a modal anchor maps cleanly onto the encoder–fusion–head path; Table 2's component ablations together with Table 4's drift diagnostics tie each constraint to a measurable drift reduction rather than asserting retention globally. This makes the central claim—that preserving acquisition-time targets plus combined feature and policy constraints lowers final action error—traceable to specific measurements.
  2. Cross-suite evidence and disciplined variant comparisons. Results span three differently structured LIBERO suites (manipulated objects, goal reaching, spatial relations). The paper includes both a fresh-target variant, which recomputes targets from its own previous-task model in the spirit of LifelongVLA rolling targets, and a raw-free variant, which drops the modal anchor, so the comparison separates persistent targets from rolling targets and cached inputs from raw observations. This level of ablation discipline is above average for robotics continual-learning submissions.
Critical weaknesses
  1. Evidence is bounded to one architecture and an offline prediction setting. All results use one controlled architecture (CLIP-LoRA-GMM) and offline prediction on LIBERO demonstrations; Goal and Spatial use only two seeds and Object three. Several lifelong VLA baselines named in Related Work (LifelongVLA, PHASER, Memory Anchors, HiDe-PET, O-LoRA, InfLoRA) are not included in the main tables under a matched architecture, optimizer, and buffer budget, so the comparative claim rests on a narrow frontier. The limitations section itself notes that closed-loop retention and transfer to other policy families, including pretrained-action-policy VLAs, remain to be established; until that evidence arrives, the generality of the persistent-anchor design is limited to this controlled offline setting.
  2. Metric ordering disagrees, and the storage–speed trade-off is quantified only in part. MSE and behavior-cloning NLL rank methods oppositely: M2Distill-R reaches lower final NLL (−12.850) than PathComp (−11.552) on Object, while PathComp has lower action MSE. The paper correctly states the two metrics do not imply one another, but does not explain why anchoring the cached mixture improves MSE without improving likelihood, or whether the MSE gain comes at the cost of compressed distribution shapes. Table 4's drift diagnosis is reported on a single seed, so its statistical uncertainty is not characterized; the storage–speed trade-off between full PathComp, which caches raw windows, and PathComp-RF is reported only in part, and the raw-free retention benefit should be weighed against this cost more explicitly.
Questions for authors
  1. Why do MSE and NLL rank methods in opposite order? Does the core policy KL term anchor cached mixture weights and compress mixture shapes, lowering coordinate error without improving log-likelihood? Could the authors report per-component NLL or entropy diagnostics to explain this gap?
  2. The confidence weight assigns anchor weight by entropy rank (Eq. 5). Under task streams with highly uneven difficulty, does this ranking systematically discard low-confidence, high-value anchors? Could the authors provide a sensitivity analysis of anchor weight distribution against forgetting performance, analogous to the raw-window ablation in Table 2?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.