Pith. sign in

REVIEW 3 major objections 4 minor 59 references

LifelongVLA claims that splitting LoRA adaptation into fast and slow pathways plus stochastic prefix replay lets a frozen vision-language-action robot policy add skills sequentially while cutting forgetting to 11.4% and replay memory by 43%

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 00:53 UTC pith:CBK2EBEN

load-bearing objection A practically useful continual-VLA recipe with an internally consistent headline result, but the evidence as printed has a table-text contradiction, no variance, and an unmeasured prefix-drift risk. the 3 major comments →

arxiv 2607.14852 v2 pith:CBK2EBEN submitted 2026-07-16 cs.RO

Towards Human-like Physical Intelligence: Lifelong Vision-Language-Action Learning for Robotic Manipulation

Yao He , Gan Sun , Wenqi Liang , Fazeng Li , Yang Cong This is my paper
classification cs.RO
keywords lifelong learningcontinual learningvision-language-action modelsrobotic manipulationcatastrophic forgettinglow-rank adaptationreplay bufferdiffusion action
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that a frozen vision-language-action robot policy can learn a stream of manipulation skills one after another without erasing earlier skills, and that the forgetting it suffers can be held below twelve percent while using far less rehearsal storage than full-trajectory replay. The proposed method, LifelongVLA, splits adaptation into a fast short-term adapter that absorbs the new task and a slow long-term adapter that consolidates old skills, with a task-aware gate mixing the two in weight space. A cached-prefix replay scheme stores only compact anchors from each skill and resamples diffusion noise at replay time to generate fresh supervision. On a ten-task stream the method reports 83.2% average success against a joint-training ceiling of 89.6%, and average forgetting of 11.4%. If the result holds, it would make continuous robot deployment practical without retraining on full histories.

Core claim

The paper's central claim is that the plasticity-stability trade-off in continuous VLA learning can be largely resolved by separating adapters by timescale while keeping the backbone frozen. Short-term LoRA parameters are driven by the new-task loss; long-term LoRA parameters are driven only by replay and distillation losses; and a gate over frozen-prefix features decides, per sample and per layer, how much of each pathway enters the weights. For old skills, the model stores only stop-gradient prefix tokens and compact action-state targets; at each replay step it draws fresh diffusion variables and recomputes the suffix with the current model, so a small cache yields many stochastic rehearsa

What carries the argument

Two modules carry the argument. Dual-Timescale LoRA Gating decomposes each low-rank weight update into a short-term plastic path and a long-term stable path, composed as (1−α)ΔW_short + αΔW_long, where α is produced by a shared gate reading frozen prefix features so the gate does not depend on its own outputs; this gives an explicit per-layer dial between new-skill plasticity and old-skill stability in a single forward pass. Cache-Efficient Stochastic Replay stores stop-gradient prefix tokens from a few random samples per skill plus compact state-action supervision, then at replay time resamples diffusion time and noise, recomputing only the suffix tokens with the current model; the cached p

Load-bearing premise

The load-bearing assumption is that the cached, stop-gradient prefix anchors faithfully represent old skills even as the LoRA weights keep changing; if those anchors drift out of alignment with the adapted model, the replay and distillation losses train toward stale targets and the reported low forgetting would not hold on longer streams.

What would settle it

Measure, after each new task, the similarity between a cached prefix token and the same sample's prefix recomputed under the current adapted model, and correlate that similarity with per-task forgetting; if cached-to-fresh similarity drops sharply on a task while its forgetting rate climbs, the replay-anchor premise fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On the reported ten-task stream, average success reaches 83.2%, only 6.4 points below the joint-training upper bound.
  • Average forgetting is 11.4%, substantially below the strongest baseline's 19.6%, with replay storage cut to 95.70 MiB per task.
  • The policy needs no task identity at test time: the gate's context is computed from frozen prefix features of the current input.
  • Real-robot five-task evaluation keeps every task above 80% success, suggesting the retention mechanism transfers beyond simulation.
  • Because inference still uses a single forward pass through gated weights, the method adds no routing latency at deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the cached-prefix anchor stays faithful under continued adaptation, the same dual-timescale split should scale to longer streams; the paper itself notes longer, randomized streams are untested.
  • Editorial inference: the same short/long adapter separation could be applied to prompt-based or prefix-based continual learning for vision-language models beyond robot control.
  • Editorial inference: a direct stress test would randomize task order and report mean and variance across seeds; the paper uses a single fixed order.
  • Editorial inference: the cache-efficiency claim would be falsifiable by measuring cosine drift between cached and freshly computed prefixes as tasks accumulate.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes LifelongVLA, a continual-learning framework for vision-language-action (VLA) robotic manipulation. It keeps a frozen VLA backbone and adapts through two LoRA pathways: a short-term adapter updated on the current task and a long-term adapter updated only on replay/distillation losses, composed by a sample-dependent gate computed from stop-gradient frozen features (§4.1). To reduce replay cost, it stores only stopped-gradient prefix tokens plus state/action supervision and recomputes suffix tokens under the current model with freshly sampled diffusion variables (§4.2). Experiments on a 10-task LIBERO stream report an average success rate of 83.2%, an average forgetting rate of 11.4%, and a replay memory reduction to 95.70 MiB per task, together with a five-task real-robot xArm evaluation. The paper argues that the method balances plasticity and stability more effectively than SFT, LwF-LoRA, ER, Info-VLA, and AtomicVLA.

Significance. If the empirical claims are reproducible, LifelongVLA would be a practically useful contribution to lifelong VLA manipulation. The dual-timescale LoRA gating design is a clean way to separate fast adaptation from stable consolidation, and the gate is explicitly constructed to avoid self-circularity by using stop-gradient frozen features in Eq. (3). The cache-efficient replay idea—storing prefix tokens rather than full trajectories—addresses a real cost bottleneck for image-rich VLA policies. The main quantitative claims are arithmetically consistent with the table averages, and the reported success rate is only 6.4 points below the joint-training upper bound. The paper also includes a real-robot evaluation, which strengthens the practical relevance. However, the current evidence is not yet sufficient: there are no seeds, variances, or confidence intervals, the key ablation in Table 3 does not isolate the effect of latent replay from the effect of dual LoRA, and the replay mechanism's reliance on cached prefix stability is unquantified.

major comments (3)
  1. [§5.1, Tables 1–2] No experimental variability is reported anywhere. Every success rate and forgetting rate appears to come from a single run, with no seeds, no variance, and no confidence intervals. Given stochastic diffusion action decoding, random replay sampling, and randomized environment initialization in LIBERO, the claimed 13.0-point SR gain and 8.2-point FOR gain over ER cannot be assessed for statistical significance. I would require at least three independent seeds with mean±std, and ideally a paired significance test, for the main tables and the ablation table.
  2. [§5.2, Table 3 rows 1–3] The ablation does not isolate the effect of the cache-efficient replay representation. Row 1, 'Ours(raw data replay)', is reported with 49.99M trainable parameters, row 2, 'Ours(single LoRA)', also with 49.99M, and row 3, 'Ours(latent replay + dual LoRA)', with 99.97M. Thus the comparison between row 1 and row 3 simultaneously changes both the replay representation and the number of LoRA pathways; there is no row for raw data replay combined with dual LoRA, nor a row for latent replay combined with single LoRA that is clearly labeled as such. The text claims 'Latent replay reduces memory from 167.62 MiB/task to 95.70 MiB/task with the same 49.99M trainable parameters, while maintaining comparable SR and lower FOR,' but the rows that share 49.99M parameters differ in both replay format and LoRA configuration. The caching benefit is load-bearing for the paper's efficiency claim and needs a
  3. [§4.2, Eqs. (7)–(9); Table 3] The replay mechanism's central assertion—that recomputing only the suffix while reusing cached prefix tokens 'reduces stale mismatch between cached and current representations'—is not supported by any measurement or bound. Eq. (7) caches stop-gradient prefix tokens from the model at task time, while Eq. (9) recomputes suffixes with the current adapted model. As the adapted weights move across tasks, the cached prefix features may drift relative to the current representation, and the replay/distillation losses in Eqs. (10)–(11) could then train against stale targets. The Table 3 ablation does not isolate prefix freshness, because raw data replay differs both in storage format and in that prefixes are recomputed from raw observations. I request a controlled experiment: compare cached-prefix replay against a variant that recomputes prefixes from stored raw observations (or periodically refr
minor comments (4)
  1. [§5, experimental setup] The experimental setup omits several details needed for reproducibility: optimizer, learning rates for the short- and long-term pathways (η_sh, η_lg), batch size, LoRA layer scope, diffusion steps, evaluation episodes, and the teacher-snapshot update schedule. Please add these.
  2. [Tables 1–2] The table entries appear garbled; several values are run together without separators (e.g., '760 7290', '22221062', '96 781094'). The averages are consistent with a plausible reconstruction, but the raw formatting must be fixed so that per-task numbers are unambiguous.
  3. [§5.3] The real-robot section reports only that success rates are 'above 80%' on every task after the five-task stream. Please provide the task-wise success rates, forgetting rates, number of evaluation episodes per task, and details of the data-collection protocol; otherwise the real-robot claim cannot be verified.
  4. [§4.1, Eq. (4)] The feature projection φ^ℓ(·) is not defined. Clarify its form (e.g., a linear layer per layer or a shared MLP) and whether it is initialized randomly or from the backbone.

Circularity Check

0 steps flagged

No significant circularity: SR/FOR are rollout metrics, the gate context is deliberately decoupled from adapted weights (Eq. 3), replay targets come from stored data or detached snapshots, and the only self-citations are background.

full rationale

Walking the claimed derivation chain: LifelongVLA's central claims are empirical—83.2% SR and 11.4% FOR on a 10-task LIBERO stream and >80% SR on a five-task xArm stream. SR and FOR are obtained from rollouts, not optimized training objectives, so no reported number equals a fitted parameter or an input quantity by construction. The one place where the paper explicitly flags a potential self-dependence, the gate context, is designed to avoid it: 'If the gate context were computed from features after gated adaptation, the gate would depend on the very weights it controls. To avoid this circular dependency, the gate context c is computed from frozen prefix features' (Sec. 4.1, Eq. 3). Replay and distillation targets are also external or detached: Eq. 7 stores stopgrad prefix tokens, Eq. 8 uses stored ground-truth actions y_j with freshly sampled diffusion variables, and Eq. 11 uses a detached teacher snapshot from task t-1. None of the targets is produced by the current model's own outputs, so the replay objective does not reduce to self-confirmation. The dual-timescale LoRA decomposition and cache-efficient replay are not announced as theorems; they are validated by ablations in Table 3 against raw replay and single LoRA, providing independent empirical support. The two self-citations ([27], [28]) appear only as introductory background and do not justify any load-bearing result or forbid alternatives. The limitations section admits the evaluation is limited in task scale/diversity and that longer streams, randomized orders, and variance are not reported; that is an external-validity concern, not circularity. No step in the paper's derivation chain was found to reduce to its own inputs.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 4 invented entities

The paper is an engineering composition over two standard tools — LoRA adaptation and replay-plus-distillation — and introduces no new theory. The load-bearing assumptions are empirical: (a) stop-gradient prefix tokens stay faithful anchors as weights adapt (Eq. 7); (b) a one-task-old detached snapshot is a good teacher (Eq. 11); (c) pooled frozen features carry task-discriminative signal for the gate (Eq. 3); (d) a 10-task fixed-order stream measures lifelong ability. The free-parameter ledger is incomplete: the ηsh/ηlg learning-rate ratio that defines the 'dual timescale' is not reported, and the meaning of M=500 is never defined — both are load-bearing for reproducing the claims.

free parameters (8)
  • short/long-term learning rates (ηsh, ηlg) = not reported (only ηlg < ηsh stated)
    Eq. 2: the entire timescale separation — the paper's core mechanism — is controlled by these two rates, yet their values are omitted. Without them the dual-timescale behavior is unspecified.
  • LoRA rank r = 16
    Set in §5; controls adapter capacity for both pathways.
  • LoRA scaling factor γ = 16
    Set in §5; an unusually large scaling factor that materially affects the gate-composed weight updates (Eq. 5).
  • M = 500
    Set in §5 but never defined in the paper (batch size? diffusion steps? mini-batch count?), so its role is unverifiable.
  • cache quota per task ct = 50 samples/task
    Set in §5; bounds replay signal; chosen without sensitivity analysis.
  • loss weights λ, β = λ=1.0, β=0.1
    Set in §5; no ablation of these weights; β controls distillation strength in Eq. 12.
  • LoRA-adapted layer scope = unspecified
    The paper says LoRA is applied to 'adapted linear layers' (§3) but not which layers of the PaliGemma prefix encoder or Gemma action decoder are adapted.
  • teacher snapshot update schedule = unspecified (F̄_{t−1} from previous task)
    Eq. 11 uses a detached snapshot from task t−1; when snapshots are taken (per task? per step?) is not stated.
axioms (7)
  • standard math Low-rank LoRA updates W = W0 + γAB approximate full fine-tuning of the frozen VLA backbone (§3, Eq. 1).
    Standard LoRA assumption, unproved in the paper but widely validated.
  • domain assumption Re-noising cached actions with fresh diffusion variables q_τ(y, ε) yields a valid retention training signal (§4.2, Eq. 8).
    Inherited from diffusion policy training; requires the denoising objective to transfer to replay instances.
  • domain assumption Stop-gradient prefix tokens from the frozen encoder remain stable anchors after LoRA weight adaptation (§4.2, Eq. 7).
    Load-bearing: the cache-efficiency claim depends on cached prefixes not going stale; never measured or bounded in the paper.
  • domain assumption A detached snapshot of the previous model is a good teacher for old skills (§4.2, Eq. 11).
    Distillation from F̄_{t−1} assumes the snapshot faithfully represents 'old behavior'; by task t the snapshot is at least one task old.
  • domain assumption Pooled frozen-prefix features provide a gate context c that separates task-specific routing (§4.1, Eq. 3).
    The task-aware gate reads only frozen features; requires the pooled prefix to contain task-discriminative signal. Plausible but unsupported.
  • domain assumption A 10-task LIBERO selection with one fixed order stands in for lifelong deployment (§5).
    Single stream, single order, no randomization; admitted in §7 as a limitation.
  • domain assumption Success rate over an unstated number of evaluation rollouts measures skill retention (§5).
    Evaluation episode count per LIBERO task is never stated, so SR/FOR point estimates are unquantifiable.
invented entities (4)
  • short-term LoRA pathway (ΔW_sh) no independent evidence
    purpose: rapidly absorb new-task skills (plasticity)
    Architecture introduced in §4.1; its benefit is evidenced only by the paper's own single-stream ablation; no external falsifiable handle.
  • long-term LoRA pathway (ΔW_lg) no independent evidence
    purpose: consolidate old skills via replay and distillation (stability)
    Trained only on replay signals, so its 'stable knowledge' is whatever the small cache provides; validated only in-house.
  • task-aware gate α(c) no independent evidence
    purpose: weight-level composition of the two pathways per sample
    Introduced in §4.1 Eq. 4; validated only in-house; no analysis of whether gate routing correlates with actual task identity.
  • stochastic replay instances (recomputed suffix tokens) no independent evidence
    purpose: generate multiple retention signals from one cached prefix
    §4.2 Eq. 9; benefit asserted, not independently verified, and the in-house ablation contradicts the prose about its effect.

pith-pipeline@v1.3.0-alltime-deepseek · 12098 in / 24683 out tokens · 202464 ms · 2026-08-02T00:53:03.612855+00:00 · methodology

0 comments
read the original abstract

Similar to the natural capabilities of humans to sequentially learn new tasks, robots with Vision-Language-Action (VLA) models should possess lifelong learning ability to learn a new task when deployed in open-world environments. However, most recently proposed lifelong learning models aim to effectively learn the current task (plasticity) or maintain high accuracy on previous tasks (stability), while the plasticity-stability trade-off remains largely unsolved in robotic manipulation models. To address this fundamental challenge, we propose a cache-efficient lifelong Vision-Language-Action learning framework for robotic manipulation (i.e., LifelongVLA), which alleviates the plasticity-stability trade-off with a dual-timescale adaptation mechanism while achieving low-cost robotic deployment with a cache-efficient replay strategy. More concretely, we propose a dual-timescale LoRA gating module to decompose VLA adaptation into two lightweight pathways: a short-term adapter for plasticity and a long-term adapter for stable consolidation. These pathways are integrated via a task-aware gate, enabling explicit control of the plasticity-stability trade-off. In the skill replay phase, a cache-efficient stochastic replay strategy is proposed to preserve more balanced retention signals without full-trajectory storage. Finally, experiments show that LifelongVLA outperforms existing baselines, demonstrating efficient skill expansion, robust retention of learned manipulation behaviors, and reduced reliance on retraining for real-world deployment on an xArm robot.

Figures

Figures reproduced from arXiv: 2607.14852 by Fazeng Li, Gan Sun, Wenqi Liang, Yang Cong, Yao He.

Figure 1
Figure 1. Figure 1: Motivation and overview. (a) Lifelong VLA learning with incremental tasks, stability–plasticity balance, and low-storage replay. (b) Performance comparison of LifelongVLA with competing baselines. retaining all previous task data and repeatedly retraining over the full history can mitigate forgetting. It incurs prohibitive memory and computational costs as the number of tasks grows. To adapt into such real… view at source ↗
Figure 2
Figure 2. Figure 2: Algorithmic diagram of LifelongVLA, which consists of two modules: (i) Dual-Timescale LoRA Gating, which composes short-term and long-term LoRA pathways with a task-aware gate to balance new-skill plasticity and old-skill stability, and (ii) Cache-Efficient Stochastic Replay, which stores compact prefix tokens and reconstructs replay suffixes to provide old-skill supervision without full-trajectory replay.… view at source ↗
Figure 3
Figure 3. Figure 3: Process-level analysis. Left: stability–plasticity trade-off. Right: incremental accuracy over tasks [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Real-world and simulator task evolution. The left panel shows real-world rollouts and the right panel shows simulator rollouts. Each row illustrates language-conditioned start-to-final-state execution. space, and simulation protocol, with task identity unavailable during testing. We report the final average success rate as SR = 1 T PT i=1 aT ,i, where at,i denotes the success rate on task i after training … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 27 linked inside Pith

  1. [1]

    Do as i can, not as i say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691, 2022

    Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    org/abs/2410.24164, 2024

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π0: A visionlanguage-action flow model for general robot control, 2024a.URL https://arxiv. org/abs/2410.24164, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Robocat: A self-improving generalist agent for robotic manipulation.arXiv preprint arXiv:2306.11706, 2023

    Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex X Lee, Maria Bauzá, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation.arXiv preprint arXiv:2306.11706, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Pith/arXiv arXiv 2022

  5. [5]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023.URL https://arxiv

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023.URL https://arxiv. org/abs/2307.15818, 2024

    Pith/arXiv arXiv 2023

  6. [6]

    Riemannian walk for incremental learning: Understanding forgetting and intransigence

    Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. InProceedings of the European conference on computer vision (ECCV), pages 532–547, 2018

    2018

  7. [7]

    On tiny episodic memories in continual learning.arXiv preprint arXiv:1902.10486, 2019

    Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc’Aurelio Ranzato. On tiny episodic memories in continual learning.arXiv preprint arXiv:1902.10486, 2019

    Pith/arXiv arXiv 1902

  8. [8]

    Diffusion policy: Visuomotor policy learning via action diffusion.arXiv preprint arXiv:2303.04137, 2023

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion.arXiv preprint arXiv:2303.04137, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Openhelix: A short survey, empirical analysis, and open-source dual-system vla model for robotic manipulation.arXiv preprint arXiv:2505.03912, 2025

    Can Cui, Pengxiang Ding, Wenxuan Song, Shuanghao Bai, Xinyang Tong, Zirui Ge, Runze Suo, Wanqi Zhou, Yang Liu, Bofang Jia, et al. Openhelix: A short survey, empirical analysis, and open-source dual-system vla model for robotic manipulation.arXiv preprint arXiv:2505.03912, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Loss of plasticity in deep continual learning.Nature, 632(8026):768–774, 2024

    Shibhansh Dohare, J Fernando Hernandez-Garcia, Qingfeng Lan, Parash Rahman, A Rupam Mahmood, and Richard S Sutton. Loss of plasticity in deep continual learning.Nature, 632(8026):768–774, 2024

    2024

  11. [11]

    Palm-e: An embodied multimodal language model

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023

    2023

  12. [12]

    Benchmarking vision, language, & action models in procedurally generated, open ended action environ- ments.arXiv preprint arXiv:2505.05540, 2025

    Pranav Guruprasad, Yangyue Wang, Sudipta Chowdhury, Harshvardhan Sikka, and Paul Pu Liang. Benchmarking vision, language, & action models in procedurally generated, open ended action environ- ments.arXiv preprint arXiv:2505.05540, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. volume 1, page 3, 2022

    2022

  14. [14]

    Voxposer: Compos- able 3d value maps for robotic manipulation with language models.arXiv preprint arXiv:2307.05973, 2023

    Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Compos- able 3d value maps for robotic manipulation with language models.arXiv preprint arXiv:2307.05973, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    5: A vision-language-action model with open-world generalization

    P Intelligence, K Black, N Brown, J Darpinian, K Dhabalia, D Driess, A Esmail, M Equi, C Finn, N Fusai, et al.π0. 5: A vision-language-action model with open-world generalization. arxiv 2025.arXiv preprint arXiv:2504.16054, 2025. 9

    Pith/arXiv arXiv 2025

  16. [16]

    Vima: General robot manipulation with multimodal prompts.arXiv preprint arXiv:2210.03094, 2(3):6, 2022

    Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts.arXiv preprint arXiv:2210.03094, 2(3):6, 2022

    Pith/arXiv arXiv 2022

  17. [17]

    Srt-h: A hierarchical framework for autonomous surgery via language-conditioned imitation learning.Science robotics, 10(104):eadt5254, 2025

    Ji Woong Kim, Juo-Tung Chen, Pascal Hansen, Lucy Xiaoyang Shi, Antony Goldenberg, Samuel Schmidgall, Paul Maria Scheikl, Anton Deguet, Brandon M White, De Ru Tsai, et al. Srt-h: A hierarchical framework for autonomous surgery via language-conditioned imitation learning.Science robotics, 10(104):eadt5254, 2025

    2025

  18. [18]

    Openvla: An open-source vision-language- action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language- action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of sciences, 114(13): 3521–3526, 2017

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of sciences, 114(13): 3521–3526, 2017

    2017

  20. [20]

    Reinforcement fine-tuning naturally mitigates forgetting in continual post-training.arXiv preprint arXiv:2507.05386, 2025

    Song Lai, Haohan Zhao, Rong Feng, Changyi Ma, Wenzhuo Liu, Hongbo Zhao, Xi Lin, Dong Yi, Qingfu Zhang, Hongbin Liu, et al. Reinforcement fine-tuning naturally mitigates forgetting in continual post-training.arXiv preprint arXiv:2507.05386, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. 2021

    2021

  22. [22]

    Remem-vla: Empowering vision-language-action model with memory via dual-level recurrent queries.arXiv preprint arXiv:2603.12942, 2026

    Hang Li, Fengyi Shen, Dong Chen, Liudi Yang, Xudong Wang, Jinkui Shi, Zhenshan Bing, Ziyuan Liu, and Alois Knoll. Remem-vla: Empowering vision-language-action model with memory via dual-level recurrent queries.arXiv preprint arXiv:2603.12942, 2026

    arXiv 2026

  23. [23]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. 2021

    2021

  24. [24]

    Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence, 40(12):2935–2947, 2017

    Zhizhong Li and Derek Hoiem. Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence, 40(12):2935–2947, 2017

    2017

  25. [25]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. volume 40, pages 2935–2947. IEEE, 2017

    2017

  26. [26]

    Code as policies: Language model programs for embodied control

    Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. 2022

    2022

  27. [27]

    Never-ending behavior-cloning agent for robotic manipulation.arXiv preprint arXiv:2403.00336, 2024

    Wenqi Liang, Gan Sun, Yao He, Yu Ren, Jiahua Dong, and Yang Cong. Never-ending behavior-cloning agent for robotic manipulation.arXiv preprint arXiv:2403.00336, 2024

    arXiv 2024

  28. [28]

    Pixelvla: Advancing pixel-level understanding in vision-language-action model.arXiv preprint arXiv:2511.01571, 2025

    Wenqi Liang, Gan Sun, Yao He, Jiahua Dong, Suyan Dai, Ivan Laptev, Salman Khan, and Yang Cong. Pixelvla: Advancing pixel-level understanding in vision-language-action model.arXiv preprint arXiv:2511.01571, 2025

    arXiv 2025

  29. [29]

    Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies.arXiv preprint arXiv:2508.20072, 2025

    Zhixuan Liang, Yizhuo Li, Tianshuo Yang, Chengyue Wu, Sitong Mao, Tian Nian, Liuao Pei, Shunbo Zhou, Xiaokang Yang, Jiangmiao Pang, et al. Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies.arXiv preprint arXiv:2508.20072, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Showui: One vision-language-action model for gui visual agent

    Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Shiwei Wu, Zechen Bai, Stan Weixian Lei, Lijuan Wang, and Mike Zheng Shou. Showui: One vision-language-action model for gui visual agent. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19498–19508, 2025

    2025

  31. [31]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023. 10

    2023

  32. [32]

    Pretrained vision-language-action models are surprisingly resistant to forgetting in continual learning.arXiv preprint arXiv:2603.03818, 2026

    Huihan Liu, Changyeon Kim, Bo Liu, Minghuan Liu, and Yuke Zhu. Pretrained vision-language-action models are surprisingly resistant to forgetting in continual learning.arXiv preprint arXiv:2603.03818, 2026

    arXiv 2026

  33. [33]

    Spatial- temporal aware visuomotor diffusion policy learning.arXiv preprint arXiv:2507.06710, 2025

    Zhenyang Liu, Yikai Wang, Kuanning Wang, Longfei Liang, Xiangyang Xue, and Yanwei Fu. Spatial- temporal aware visuomotor diffusion policy learning.arXiv preprint arXiv:2507.06710, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Packnet: Adding multiple tasks to a single network by iterative pruning

    Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765–7773, 2018

    2018

  35. [35]

    Preserving and combining knowledge in robotic lifelong reinforcement learning.Nature Machine Intelligence, 7(2):256–269, 2025

    Yuan Meng, Zhenshan Bing, Xiangtong Yao, Kejia Chen, Kai Huang, Yang Gao, Fuchun Sun, and Alois Knoll. Preserving and combining knowledge in robotic lifelong reinforcement learning.Nature Machine Intelligence, 7(2):256–269, 2025

    2025

  36. [36]

    Llarva: Vision-action instruction tuning enhances robot learning.arXiv preprint arXiv:2406.11815, 2024

    Dantong Niu, Yuvan Sharma, Giscard Biamby, Jerome Quenum, Yutong Bai, Baifeng Shi, Trevor Darrell, and Roei Herzig. Llarva: Vision-action instruction tuning enhances robot learning.arXiv preprint arXiv:2406.11815, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Open x-embodiment: Robotic learning datasets and rt-x models.arXiv preprint arXiv:2310.08864, 2023

    Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models.arXiv preprint arXiv:2310.08864, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, 2025

    Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Physical Intelligence, Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokinsky, Shihao Cao, Thomas Charbonnier, et al.π0.7: A Steerable Generalist Robotic Foundation Model with Emergent Capabilities.arXiv preprint arXiv:2604.15483, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, Christoph H Lampert, et al. Incremental classifier and representation learning. InConference on Computer Vision and Pattern Recognition (CVPR), pages 5533–5542, 2024

    2024

  41. [41]

    A generalist agent.arXiv preprint arXiv:2205.06175, 2022

    Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent.arXiv preprint arXiv:2205.06175, 2022

    Pith/arXiv arXiv 2022

  42. [42]

    Progressive neural networks.arXiv preprint arXiv:1606.04671, 2016

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Ko- ray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks.arXiv preprint arXiv:1606.04671, 2016

    Pith/arXiv arXiv 2016

  43. [43]

    Rl’s razor: Why online reinforcement learning forgets less.arXiv preprint arXiv:2509.04259, 2025

    Idan Shenfeld, Jyothish Pari, and Pulkit Agrawal. Rl’s razor: Why online reinforcement learning forgets less.arXiv preprint arXiv:2509.04259, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Cliport: What and where pathways for robotic manipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Cliport: What and where pathways for robotic manipulation. 2021

    2021

  45. [45]

    Perceiver-actor: A multi-task transformer for robotic manipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. InConference on Robot Learning, pages 785–799. PMLR, 2023

    2023

  46. [46]

    Smolvla: A vision-language- action model for affordable and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language- action model for affordable and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Coda-prompt: Continual decomposed attention- based prompting for rehearsal-free continual learning

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention- based prompting for rehearsal-free continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11909–11919, 2023. 11

    2023

  48. [48]

    Continual learning in vision-language models via aligned model merging.arXiv preprint arXiv:2506.03189, 2025

    Ghada Sokar, Gintare Karolina Dziugaite, Anurag Arnab, Ahmet Iscen, Pablo Samuel Castro, and Cordelia Schmid. Continual learning in vision-language models via aligned model merging.arXiv preprint arXiv:2506.03189, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    A comprehensive survey of continual learning: Theory, method and application.IEEE transactions on pattern analysis and machine intelligence, 46(8): 5362–5383, 2024

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application.IEEE transactions on pattern analysis and machine intelligence, 46(8): 5362–5383, 2024

    2024

  51. [51]

    Dualprompt: Complementary prompting for rehearsal-free continual learning.arXiv preprint arXiv:2204.04799, 2022

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning.arXiv preprint arXiv:2204.04799, 2022

    Pith/arXiv arXiv 2022

  52. [52]

    Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation.IEEE Robotics and Automation Letters, 2025

    Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Zhibin Tang, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation.IEEE Robotics and Automation Letters, 2025

    2025

  53. [53]

    Continual world: A robotic benchmark for continual reinforcement learning

    Maciej Wołczyk, Michał Zając, Razvan Pascanu, Łukasz Kuciński, and Piotr Miłoś. Continual world: A robotic benchmark for continual reinforcement learning. volume 34, pages 28496–28510, 2021

    2021

  54. [54]

    Long-horizon language-conditioned imitation learning for robotic manipulation

    Xiangtong Yao, Tobias Blei, Yuan Meng, Yu Zhang, Hongkuan Zhou, Zhenshan Bing, Kai Huang, Fuchun Sun, and Alois Knoll. Long-horizon language-conditioned imitation learning for robotic manipulation. IEEE/ASME Transactions on Mechatronics, 2025

    2025

  55. [55]

    Boosting continual learning of vision-language models via mixture-of-experts adapters

    Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219–23230, 2024

    2024

  56. [56]

    Atomicvla: Unlocking the potential of atomic skill learning in robots.arXiv preprint arXiv:2603.07648, 2026

    Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, Hang Xu, Hefeng Wu, et al. Atomicvla: Unlocking the potential of atomic skill learning in robots.arXiv preprint arXiv:2603.07648, 2026

    arXiv 2026

  57. [57]

    Mllm-cl: Continual learning for multimodal large language models.arXiv preprint arXiv:2506.05453, 2025

    Hongbo Zhao, Fei Zhu, Haiyang Guo, Meng Wang, Rundong Wang, Gaofeng Meng, and Zhaoxiang Zhang. Mllm-cl: Continual learning for multimodal large language models.arXiv preprint arXiv:2506.05453, 2025

    arXiv 2025

  58. [58]

    Information-theoretic constraints for continual vision-language-action alignment.arXiv preprint arXiv:2603.13335, 2026

    Libang Zhao, Qixin Zeng, Hongyin Zhang, and Donglin Wang. Information-theoretic constraints for continual vision-language-action alignment.arXiv preprint arXiv:2603.13335, 2026

    arXiv 2026

  59. [59]

    imanip: Skill-incremental learning for robotic manipulation

    Zexin Zheng, Jia-Feng Cai, Xiao-Ming Wu, Yi-Lin Wei, Yu-Ming Tang, Ancong Wu, and Wei-Shi Zheng. imanip: Skill-incremental learning for robotic manipulation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13890–13900, 2025. 12

    2025