Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning

Summary

This paper asks whether the continual-learning conclusions that dominate the robot-imitation literature — pervasive catastrophic forgetting, the need for large replay buffers or careful regularization to combat it, and a stark stability–plasticity trade-off — still hold when the underlying policy is a modern pretrained Vision-Language-Action model rather than a small BC network trained from scratch. The authors run a systematic experience-replay sweep on the four LIBERO suites with Pi0 and GR00T N1.5 against BC-Transformer / BC-ViT / BC-Diffusion baselines; they find that at the standard LIBERO buffer (20% of task data) both VLAs achieve NBT $\approx$ 0 and sometimes slight positive backward transfer, while the small BC policies exhibit substantial forgetting; that the gap widens as the buffer shrinks; that a pretraining ablation (Pi0 from VL+Action vs. from VL vs. from scratch) attributes much of the forgetting resistance to pretraining rather than the architecture; that a component-swap analysis attributes most of the residual forgetting to the VL backbone; and that finetuning restores previously "forgotten" tasks in fewer than $\sim$10% of the original training steps, whereas BC-Transformer often needs comparable or more time. A small real-robot Pi0 pilot on three Mutex tasks corroborates the qualitative picture. The conceptual move — treating pretrained VLAs as a distinct continual-learning regime governed by representation reuse rather than by specialized CL algorithms — is timely and, if the empirical evidence holds up, would meaningfully redirect near-term work on lifelong robot learning. The paper's main conceptual limitation is that its headline explanation (pretraining) is confounded with model scale and, in the GR00T case, with finetuning scope: Appendix C already shows that a matched-size from-scratch Pi0 approaches the same NBT regime, and GR00T's frozen VL backbone means the very mechanism the paper isolates in Section 6 is not actually exercised in one of the two headline models. The empirical breadth (single task ordering, LIBERO-only main results, three-task real-robot pilot) is also thinner than the sweeping "distinct continual-learning regime" framing suggests, and the code trail behind the quantitative results turns out to be split across sibling repos that the paper does not link.

Strengths

• Timely and well-motivated question. The community has largely inherited its continual-learning intuitions from small-network settings; asking whether they transfer to pretrained VLAs is exactly the right question, and the paper frames it clearly in Section 1.
• Multi-axis empirical protocol. The combination of (i) buffer-size sweep, (ii) non-ER baselines (Sequential, EWC), (iii) pretraining ablation with three Pi0 initializations, (iv) component-swap probes, and (v) recovery-efficiency analysis is a substantially more thorough package than a single headline table would have been.
• Metric hygiene. Introducing a normalized NBT in Appendix A.2 to defuse the "high-SR models look worse under absolute NBT" artifact is a useful methodological contribution in its own right.
• Recovery finding is genuinely surprising. The Table 5 / Figure 8 result that Pi0 recovers "forgotten" tasks in a small fraction of the original training budget while BC-Transformer often needs comparable time is a clean and interpretable signal that pretrained representations retain more than the task-level SR curve indicates.
• Honest limitations paragraph. The Limitations section on LIBERO-only evaluation, single Franka embodiment, small real-robot pilot, and Pi0-focused mechanistic studies flags most of the obvious external-validity concerns rather than hiding them.

Weaknesses

• Pretraining vs. scale confound in the headline framing. The main-body comparison sets Pi0 / GR00T ($\sim$3B) against BC baselines ($\sim$15–26M) and attributes the gap to pretraining. But Appendix C, Table 11 shows that a 250M-LLM + 400M-SigLIP Pi0 trained from scratch already reaches NBT $\approx -0.05$ — within the "resistant to forgetting" regime. This directly weakens the causal story of Section 5 and the framing of Section 1, and is buried in a paragraph labelled "initial investigation."
• GR00T's frozen VL is a confound for the Section-6 mechanism. Table 9 shows GR00T freezes both the vision encoder and the language model during continual learning; Section 6 concludes the VL backbone is the dominant source of forgetting. Under the paper's own mechanism, GR00T's near-zero NBT could be explained trivially by the fact that its VL is not updated at all. Neither of the natural controls (GR00T with unfrozen VL, or Pi0 with frozen VL) is reported.
• "Zero forgetting with small replay" overclaim. The abstract and introduction advertise "zero forgetting even with a small replay data size (e.g., 2%)," but Section 4.2 itself states that at the 2% buffer VLAs have NBT $\approx$ 0.1–0.2. The near-zero results are at the 20% buffer used in Table 1. This is the single most repeated headline claim and it materially overstates what the body demonstrates.
• Generalization from "Pi0 from scratch" to "non-pretrained models." Section 5 argues that "low measured forgetting in non-pretrained models can arise from an inability to learn," using Figure 5's KT curves. But those curves only compare three Pi0 variants; the paper's actual non-pretrained models (BC-Transformer et al.) have high, not low, NBT (Table 4). The mechanism applies to Pi0-from-scratch specifically, not to the general "non-pretrained" category the sentence claims.
• Single task ordering. Every experiment uses one fixed task ordering per suite; continual-learning outcomes are known to be sensitive to order, and this variance is neither bounded nor reported. Combined with unstated seed counts, order-level noise is unaccounted for.
• "Diverse VLA backbones" is only two models with similar recipes. Pi0 and GR00T both use a SigLIP vision encoder plus a Gemma/Qwen-style LLM plus a flow-matching action head. The "diverse" framing is stronger than the sample supports; OpenVLA / RT-2 / Octo / discretized-action VLAs cited in Related Work are absent.
• Compartmentalization is one reading of the swap experiment. The Section-6 swaps show both $\mathrm{VL}_{k+1}$ and $A_{k+1}$ hurt performance vs. the original, with VL hurting more. This is consistent with compartmentalized knowledge but also with a joint-representation view where VL simply carries more task capacity. The paper commits to the former without a control that distinguishes them (e.g., late-layer vs. early-layer VL swaps).
• Mechanistic studies rest on a single VLA. Pretraining ablations, component swaps, and recovery analyses are all Pi0-only, even though the manuscript's headline is written across "VLAs" plural. GR00T is precisely the model where the mechanism predicts something different (frozen VL), so this omission is load-bearing rather than incidental.
• Real-robot pilot is undersized for the role it plays in the narrative. Three tasks, single Pi0 run, no baselines, no seed information; the pilot is fine as a sanity check but is currently framed as evidence that the finding transfers to real robots. (The reported NBT = $-0.003$ is internally consistent with Table 3 — I re-derived it — but $K=3$, $N=1$ cannot support the qualitative-generalization framing without a Sequential or EWC contrast on the same setup.)
• Missing display equations. The continual-learning objective (Section 3.1), the standard NBT (Section 4.1), and the normalized NBT (Appendix A.2) are all introduced with a colon and no visible formula. Since the paper's sign convention is opposite to Lopez-Paz & Ranzato's BWT, the missing NBT formula is load-bearing for every reported number.
• Recovery-efficiency claim slightly overstated. "Fewer than 10% of the original training steps ... across all benchmarks" is falsified by the 0.105 (10.5%) entry for LIBERO-10 in Table 5. Minor but factual.
• Section-3 / Table-2 pointer errors. The intro pointer to Section 3 for the forgetting-resistance results should be Section 4; the "neither baseline consistently improves over nonVLA policies" sentence references Table 2, but Table 2 contains only Pi0 and GR00T rows (no non-VLA comparison).
• Objective-ablation summary understates the NBT movement. Switching flow matching to $\ell_2$ moves NBT from $-0.0003$ to $+0.016$ — a sign flip and >50$\times$ magnitude change on the metric the paper cares most about. Calling this "little difference" is fine for SR but glosses over the NBT delta.
• Notational discontinuity between Section 3.1 and Section 3.2. The CL objective conditions on $(s, \mathcal{T})$; the VLA definition switches to $(l, \mathbf{o}_{\leq t})$. A single bridging sentence would reconcile them.
• NBT sign-convention readability trap. "Positive backward transfer" (good) = "negative NBT" (good); "positive NBT" is bad. Table 2 bolds the best (most negative) NBT values without a caption reminder; a one-line sign-convention pointer near each NBT-bearing table would prevent repeated re-reading.

Reproducibility & code

The veritas execution attempt could not independently reproduce any of the paper's quantitative claims: the reproduction environment was pinned to a single CPU core, which made full-scale Pi0 / GR00T continual-learning training and MuJoCo closed-loop evaluation infeasible. That is an environmental limitation of the run, not a fault of the paper. However, the same attempt did surface concrete availability and specification gaps that would obstruct any lab attempting full-scale reproduction:

• Project "website" contains no code. The URL advertised in the abstract points to a website-only repository (index.html + static/). The runnable implementation is spread across three sibling GitHub repos (continual-openpi, continual-GR00T, and a fork of libero) that the paper never names. This was only discovered by inspection; it should be linked from the paper.
• Pretraining-ablation checkpoints are not released. Section 5's causal story rests on three checkpoints — Pi0 from VL+Action, Pi0 from VL, Pi0 from scratch — of which only the first is publicly available. The other two are credited to individuals in the Acknowledgments, with no download link, HuggingFace repo, or full pretraining recipe (data mixture, steps, LR schedule, hardware). Independent verification therefore requires re-doing large-scale VLA pretraining.
• Number of seeds unstated. Every $\pm$std in Tables 1, 2, 4, 6, 7, 8 is anonymous with respect to seed count; the aggregation script does not fix a canonical value. Confidence-interval reproduction and implicit significance claims are under-specified.
• EWC hyperparameters missing. The openpi.continual.ewc implementation exposes $\lambda$, $\gamma$, and Fisher-batch count; Appendix B.5 defines EWC only in prose. Because Table 2's "EWC does not help pretrained VLAs" conclusion depends critically on $\lambda$, these values should be tabulated (ideally with a small $\lambda$ sweep).
• Recovery protocol under-specified. For Table 5 / Figure 8, the starting checkpoint (which post-Task-$k+1$ model at which buffer size), the peak-SR detection rule, the eval interval, and the handling of runs that never regain the original peak are not stated.
• Model-size ablation configs missing. Table 11's 17M and 250M LLM+action-expert variants are not specified beyond header labels; n_layers, d_model, n_heads, and per-variant training budgets are not tabulated.
• Real-robot pilot is not reproducible from the paper. Beyond a physical Franka + Mutex kitchen, the exact demonstrations, per-task success criterion, rollout counts, and evaluation code path are not released, and Table 3 lacks trial information.
• Released code required ~13 patches to install and start. Most were dependency/environment drift (an LEROBOT_HOME $\rightarrow$ HF_LEROBOT_HOME env-var rename, a datasets==5.0.0 / lerobot conflict, torch.load(weights_only=...) and np.bool compatibility shims, fork-after-CUDA-init requiring spawn for LIBERO GPU workers). Two, though, are in the authors' own code (examples/libero/convert_libero_data_to_lerobot.py REPO_NAME/mkdir mismatch; Gr00tPolicy._load_metadata writing metadata_content: None as a type hint rather than a default), which indicates those specific paths were not exercised end-to-end before release. None of these fixes touch methodology, but they slow down any reproducer.

Recommended Changes

Essential

• Disentangle pretraining from model scale in the main narrative. Promote Appendix C, Table 11 into Section 5 (or a new section) so the reader sees that a matched-scale from-scratch Pi0 already achieves NBT $\approx -0.05$. Tighten the causal claim from "pretraining plays an integral role" to something like "pretraining and model scale jointly explain most of the forgetting resistance."
• Add a control that isolates the VL-backbone finetuning-scope confound. Either run GR00T with unfrozen VL or Pi0 with frozen VL on at least one suite so that the Section-6 mechanism ("VL backbone is the dominant source of forgetting") is not entangled with GR00T's frozen-VL finetuning strategy.
• Rewrite the "zero forgetting with small replay" framing. In both the abstract and Section 1, separate the near-zero regime (20% buffer, Table 1) from the small-buffer regime (2%, NBT $\approx$ 0.1–0.2, Section 4.2). Remove the "e.g., 2% of training data" clause from the zero-forgetting claim.
• Show the CL-objective, NBT, and normalized-NBT equations explicitly where each is introduced (Sections 3.1, 4.1, and Appendix A.2), including the handling of $c_{k,k}=0$ for the normalized variant. Given the non-standard sign convention, the NBT display in particular is load-bearing.
• Release the "Pi0 from VL" and "Pi0 from scratch" checkpoints used in Section 5, or provide a complete pretraining recipe (data mixture, steps, LR schedule, hardware) sufficient to reconstruct them.
• Link the actual code repositories from the paper. The "project website" pointer currently reaches only a static site; name continual-openpi, continual-GR00T, and the LIBERO fork (or equivalents) explicitly, and state which experiments each contains.
• Report the number of seeds per configuration for every $\pm$std number in Tables 1, 2, 4, 6, 7, 8, and add the EWC hyperparameters ($\lambda$, $\gamma$, Fisher batches) used for Table 2.

Suggested

• Add a small task-ordering sweep on one suite (e.g., 3–5 orderings on LIBERO-Object) to bound the order-level variance the current protocol ignores.
• Correct the recovery-efficiency claim from "fewer than 10%" to "$\lesssim$ 10%" or flag LIBERO-10 as the exception.
• Fix the Section-3 / Table-2 pointer issues: the intro reference for the forgetting-resistance result should be Section 4, and the "neither baseline improves over nonVLA policies" sentence should be reworded (e.g., "neither Sequential nor EWC reduces forgetting as well as ER does for pretrained VLAs") or moved next to Table 1 / Figure 2.
• Scope the "non-pretrained models" generalization in Section 5 to the Pi0-from-scratch variant specifically, since BC-Transformer does not exhibit low measured forgetting.
• Reconcile the Figure-6(c) description with the Figure-8 experiment. Either add the component-swap recovery experiment that Figure 6(c) advertises, or reframe Figure 6(c) so the recovery panel matches the whole-model finetune that is actually reported.
• Fix the Table 9 caption so "LoRA adapters ... frozen backbone" is scoped to language + action (the vision encoder is fully fine-tuned per the strategy row).
• Add a "diverse VLA backbones" qualifier: acknowledge that Pi0 and GR00T share the SigLIP + Gemma-family + flow-matching recipe, and either add one architecturally different VLA (e.g., discretized-action) or downgrade the framing to "two representative VLAs."
• Fully specify the recovery experiment protocol (starting checkpoint, peak-SR detection rule, eval interval, handling of non-recovery) so Table 5 / Figure 8 are unambiguous, and document the Table-11 model-size variants (n_layers, d_model, whether Table 10's schedule was reused).
• Ship a working environment (pinned environment.yml / Dockerfile / release-checklist run). The install-time patches surfaced during reproduction — including two in the authors' own code paths — indicate the release was not exercised end-to-end in a clean environment.
• Release the real-robot evaluation script and per-task rollout counts so the Mutex pilot (Table 3) is reproducible by any lab with the same hardware, and add a Sequential or EWC contrast on the same three tasks.
• Reword the objective-ablation summary (Table 12) so the sign flip in NBT is acknowledged rather than characterized as "little difference."
• Add a one-line sign-convention reminder near every NBT-bearing table ("NBT $\uparrow$ = more forgetting; negative NBT = positive backward transfer") to defuse the persistent double-negative readability trap.