Pith. sign in

REVIEW 3 major objections 4 minor 56 references

A robot policy can keep minute-long visual history by turning each past camera frame into one token from its own vision encoder, lifting success from about 32% to 84% without new memory modules.

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 · grok-4.5

2026-07-10 23:33 UTC pith:MFESFH7G

load-bearing objection Solid systems paper: one native token per past frame-view, action-supervised under a frozen VLA, with large sim/real gains and clean efficiency numbers; residual risk that Stage-1 under-specifies abstract progress on a narrow suite. the 3 major comments →

arxiv 2607.06678 v1 pith:MFESFH7G submitted 2026-07-07 cs.RO

NativeMEM: Native Memory Compression for Long-Horizon Robotic Manipulation

Ziye Wang , Modi Shi , Chaojun Ni , Jiazhi Yang , Mengdi Li , Zhizhong Su , Tianwei Lin , Hongyang Li This is my paper
classification cs.RO
keywords Vision-Language-ActionNative Memory Compressionlong-horizon manipulationmemory tokenizerrobotic memorydata efficiencyreal-time inference
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.

Pretrained vision-language-action policies usually act on the current camera frame only, so they fail when the right move depends on what happened earlier—which buttons were already pressed, where a block started, or which grocery items were scanned. This paper shows that the policy’s own vision encoder can be reused to squeeze every historical frame-view into a single “native” memory token. Those tokens are trained first under a frozen policy so they live in the same visual language the model already understands, then the full policy is fine-tuned with the compact history attached. The result is long, high-frequency memory that needs no external planner or new architecture, raises average success from 32.4% to 84% in simulation and to 98.7% on real robots, and matches earlier methods while using only about 20% of the training data.

Core claim

Native Memory Compression—compressing each past frame-view into one token produced by the VLA’s own vision encoder and trained under the frozen policy’s action loss—lets a pretrained single-frame VLA attend over minute-scale visual history with negligible latency, converting it into a memory-augmented policy that substantially outperforms both VLM-driven and external-module baselines.

What carries the argument

Native Memory Compression: a memory tokenizer derived from the pretrained vision encoder that, via a learnable query token and linear projection, maps each historical frame-view to one action-supervised summary token; a two-stage pipeline freezes the VLA while learning the tokenizer, then freezes the tokenizer while fine-tuning the policy with the cached tokens appended to its input sequence.

Load-bearing premise

That the ordinary action-prediction loss of a frozen policy is enough supervision to force each single-token summary to keep the fine-grained historical details (counts, original pads, scanned items) that later actions actually need.

What would settle it

On a held-out long-horizon task whose critical cue is a subtle visual detail early in the episode, measure whether success collapses when that cue is removed or scrambled after the tokenizer has already produced its single tokens; if performance stays high, the compression is not truly retaining the claimed information.

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

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. NativeMEM converts a pretrained single-frame VLA into a memory-augmented policy by Native Memory Compression: the VLA’s own vision encoder is repurposed (with a learnable query and projection) to map each historical frame–view into one native memory token (Eqs. 5–7), which is appended to the original token sequence. Stage 1 freezes the VLA and trains the tokenizer with the native action loss on mixed standard and memory-demanding data (Eq. 9); Stage 2 freezes the tokenizer and finetunes the VLA on task demos with cached tokens (Eq. 10). On simulation (RMBench + button tasks) average success rises from 32.4% (Mem-0) to 84.0%; on real robots (including an unseen checkout task) overall success reaches 98.7%, with low latency/GPU cost and competitive results using ~20% of prior training data. Ablations (Tab. 3) and attention maps (Figs. 5–6) support the design choices.

Significance. If the result holds, the paper offers a practical, architecture-light path to long-horizon memory for pretrained VLAs: one-token-per-frame compression that reuses the existing visual pathway and attention stack, avoiding external planners or freshly initialized memory modules. The two-stage pipeline, strong sim/real gains, inference-efficiency curves (Fig. 4a), data-efficiency results (Fig. 4b), and qualitative attention evidence are concrete contributions that the community can build on. The work is empirical rather than theoretical; its value is the demonstrated conversion recipe and the efficiency–horizon trade-off it unlocks for non-Markovian manipulation.

major comments (3)
  1. §3.2 Eqs. 5–9 and Limitations: the central claim that action loss on a frozen VLA is sufficient to force each frame–view into a single token that retains fine-grained non-Markovian cues (counts, original pads, scanned items) is only partially stress-tested. Tab. 3 shows Stage-1 and long/dense history matter, and Figs. 5–6 are consistent with task-relevant focus, but there is no controlled probe that the one-token summary preserves the specific detail required rather than a coarse progress proxy that works on this suite. A short diagnostic (e.g., recovery of button-press counts or original-pad identity from memory tokens alone, or a held-out task that forces finer discrimination) would make the compression claim load-bearing rather than inferred from end-task success.
  2. Tables 1–2: success rates are reported as point estimates with no trial counts, standard errors, or confidence intervals. Given the large jumps (32.4%→84.0% sim; ~98.7% real) and the fact that several baselines are reproduced (HAMLET*, MEM-short*) on the same backbone and limited demos, uncertainty quantification is needed to support the ranking and the claim of consistent outperformance.
  3. §4.1 Baselines and Tab. 1–2: fairness of the reproduced external-memory baselines is asserted but under-specified. MEM-short is noted as originally designed for large-scale retraining; training it only on the same limited demos as NativeMEM may understate that line of work. A short sensitivity note (or an upper-bound run with more data where available) would clarify how much of the gap is method vs. training regime.
minor comments (4)
  1. Typo consistency: “Externel Memory Modules” in Tab. 1 header; “NATIVEMEM” vs “NativeMEM” capitalization varies across abstract, title, and body.
  2. Fig. 4(a) axis units and exact history lengths used for the 100 ms / 32 GB operating points should be stated in the caption or text so the efficiency claim is reproducible.
  3. §3.2: the composition of I_t (first frame + recent window + current) is described qualitatively; giving the default window length and sampling rate used in Stage 1 would help replication.
  4. Related Work: a brief explicit contrast with concurrent compressed-history VLAs (beyond HAMLET/MEM) would situate the “native” claim more sharply.

Circularity Check

0 steps flagged

No circularity: standard two-stage supervised finetuning of a memory tokenizer under frozen-VLA action loss, evaluated on held-out task success.

full rationale

NativeMEM is an empirical systems/method paper, not a first-principles derivation. The load-bearing chain is: (i) define one-token-per-frame memory via the VLA vision encoder plus a learnable query (Eqs. 5–7); (ii) train that tokenizer with the frozen VLA’s native action loss on mixed data (Eq. 9); (iii) cache tokens and finetune the VLA on target demos (Eq. 10); (iv) report success rates vs. baselines on simulation and real-robot memory tasks (Tabs. 1–2), plus ablations (Tab. 3) and efficiency plots. None of these steps reduces a claimed prediction to a fitted constant or to a definition of the target. Action-supervised tokens are optimized and then measured on independent task success; that is ordinary supervised learning, not self-definitional circularity. Reproducing HAMLET*/MEM-short* on the same π0.5 backbone and demo budget is experimental control, not a self-citation uniqueness theorem. No uniqueness result is imported from overlapping authors to force the design. The Limitations section’s own caveat about action-loss under-specifying abstract memory is a correctness/generalization risk, not circularity. Score 0 with empty steps is the honest finding.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on empirical robotics assumptions and a few design choices, not on free physical constants. Load-bearing premises are that manipulation is non-Markovian, that the pretrained VLA’s visual token space can host history, and that action loss alone teaches useful single-token summaries. Free parameters are engineering choices (token budget M=1, history sampling, update rates, data mix) that affect reported success but are not fitted “constants of nature.” Invented entities are the memory tokenizer machinery and native memory tokens themselves.

free parameters (4)
  • Per-frame memory token budget M = 1
    Set to the extreme M=1 by design (Eq. 4 discussion); controls the fidelity–efficiency trade-off and is not derived from data uniqueness.
  • History index set I_t (first frame + recent window + current) = first frame + recent window + current
    Training-time sampling of which frames enter memory is a hand-chosen scheme that shapes what the tokenizer learns; not uniquely determined.
  • Memory update frequency / horizon length at inference = dense updates; minute-level horizon claimed
    Ablations use 0.5 Hz and 2–6 s horizons; deployment claims minute-level / ~160+ frames. These are operational knobs that strongly affect success (Table 3).
  • Stage-1 data mixture (standard + memory-demanding, sim + real)
    Composition and weighting of mixed training data for the generic tokenizer are chosen by authors and affect transfer (e.g., grocery “unseen” tokenizer behavior).
axioms (4)
  • domain assumption Long-horizon manipulation is non-Markovian: identical current (o_t, s_t, ℓ) can require different optimal actions given different histories H_t (Eq. 2).
    Stated in §3.1 as the problem foundation; standard POMDP-style robotics assumption, not proved here.
  • ad hoc to paper Action prediction loss of a frozen pretrained VLA is sufficient supervision to learn compact, action-relevant single-token frame summaries (Eq. 9).
    Core training principle of Stage 1; Limitations admit a possible semantic gap for more abstract memory needs.
  • domain assumption Memory tokens produced by a vision encoder initialized from the VLA remain interpretable by that VLA’s attention when appended to the original token sequence (Eqs. 5–8).
    Compatibility claim that justifies “native” rather than external modules; supported empirically but not independently proved.
  • ad hoc to paper Reproduced external-memory baselines (HAMLET*, MEM-short*) on π0.5 with the same limited demos are fair comparators for the reported gains.
    §4.1 notes original methods may need large-scale retrain; fairness of constrained reproduction is an experimental assumption.
invented entities (2)
  • Native memory token m^v_τ (one summary per frame-view) no independent evidence
    purpose: Ultra-compact historical visual evidence injected into the pretrained VLA sequence without new memory architectures.
    Defined via encoder query + linear projection (Eqs. 5–6); existence as a useful action-aligned unit is established only inside this paper’s training loop.
  • Memory tokenizer (E_mem, q_mem, W_mem, b_mem) derived from the VLA vision encoder no independent evidence
    purpose: Learn Stage-1 action-supervised compression reusable for Stage-2 task finetuning and real-time memory queues.
    New module composition relative to the base VLA; no external falsifiable prediction beyond task success in this work.

pith-pipeline@v1.1.0-grok45 · 17843 in / 3631 out tokens · 42894 ms · 2026-07-10T23:33:29.417296+00:00 · methodology

0 comments
read the original abstract

How can pretrained Vision-Language-Action (VLA) models retain long-horizon visual histories with high-frequency updates without sacrificing efficiency? Existing approaches rely on external memory management, which restrains either the memory horizon or the reactiveness of pretrained policies. To this end, we present NativeMEM, a VLA policy that features long-term and real-time updated memory. At its core is an efficient memory encoding scheme, Native Memory Compression, which repurposes the VLA's own vision encoder to compress each historical frame from each camera view into a single token. Appended to the input sequence, these memory tokens enable the pretrained VLA to attend over long-term history with negligible latency overhead, requiring neither an external planner nor a freshly initialized memory module. To align the memory tokens with the pretrained policy, we first develop a generic memory tokenizer under the supervision of a frozen VLA on memory-demanding data, and then unfreeze the VLA for task-specific fine-tuning. NativeMEM consistently outperforms prior methods, boosting success rates from 32.4% to 84.0% in simulation and up to 98.7% on real robots, while maintaining low inference latency and GPU memory usage. Notably, NativeMEM exhibits high data efficiency by achieving competitive results with prior arts using only 20% of the training data.

Figures

Figures reproduced from arXiv: 2607.06678 by Chaojun Ni, Hongyang Li, Jiazhi Yang, Mengdi Li, Modi Shi, Tianwei Lin, Zhizhong Su, Ziye Wang.

Figure 1
Figure 1. Figure 1: NATIVEMEM differs from prior memory-augmented VLAs that rely on VLM-generated textual notes or external memory modules. (a) By repurposing the VLA’s own vision encoder, it compresses each historical frame-view observation into a single native memory token, allowing the policy to condition on the full visual history through its original token sequence. (b) This ultra￾compact representation enables minute-le… view at source ↗
Figure 2
Figure 2. Figure 2: Training pipeline of NATIVEMEM. Stage 1: We freeze the pretrained VLA and learn a native memory tokenizer, initialized from its visual encoder, using the VLA’s original action predic￾tion loss on mixed standard and memory-dependent manipulation data. The tokenizer compresses each frame-view observation into a single action-aligned memory token. Stage 2: We cache mem￾ory tokens and finetune the VLA with the… view at source ↗
Figure 3
Figure 3. Figure 3: Memory-dependent manipulation tasks. In Click Buttons, the robot must follow a specified button sequence without repeating completed presses. In Put Back Block, it must remember the block’s original pad after moving it to the center. In Grocery Checkout Scanning, it must track which items have already been scanned and avoid duplicate or missed scans. tokens at a specified update frequency and maintains a c… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Inference latency and memory consumption. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of attention from action tokens to historical memory snapshots. The x [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the spatial attention used by the memory tokenizer when compressing each frame￾view observation into a single memory token. The top row shows the tokenizer attention maps over input im￾ages, while the bottom row shows the corresponding raw observations. To understand what is encoded into each mem￾ory token, we visualize the tokenizer’s spa￾tial attention over historical observations. As sh… 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

56 extracted references · 56 canonical work pages · 23 internal anchors

  1. [1]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, 2023

    work page 2023

  2. [2]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. OpenVLA: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    $\pi_{0.5}$: 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.π 0.5: a vision-language-action model with Open-World generalization. arXiv preprint arXiv:2504.16054, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Q. Bu, Y . Yang, J. Cai, S. Gao, G. Ren, M. Yao, P. Luo, and H. Li. UniVLA: Learning to act anywhere with task-centric latent actions.arXiv preprint arXiv:2505.06111, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    G. Team, A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, J. Li, J. Zhu, L. Feng, et al. GigaBrain-0: A world model-powered vision-language-action model.arXiv preprint arXiv:2510.19430, 2025

    work page arXiv 2025

  7. [7]

    C. Ni, C. Chen, X. Wang, Z. Zhu, W. Zheng, B. Wang, T. Chen, G. Zhao, H. Li, Z. Dong, et al. SwiftVLA: Unlocking spatiotemporal dynamics for lightweight vla models at minimal overhead.CVPR 2026, 2026

    work page 2026

  8. [8]

    A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, H. Li, J. Li, J. Lv, J. Liu, et al. GigaWorld- Policy: An efficient action-centered world–action model.arXiv preprint arXiv:2603.17240, 2026

    work page arXiv 2026

  9. [9]

    H. Jang, S. Yu, H. Kwon, H. Jeon, Y . Seo, and J. Shin. ContextVLA: Vision-Language-Action model with amortized multi-frame context.arXiv preprint arXiv:2510.04246, 2025

    work page arXiv 2025

  10. [10]

    Zhang, Y

    J. Zhang, Y . Chen, Y . Xu, Z. Huang, Y . Zhou, Y .-J. Yuan, X. Cai, G. Huang, X. Quan, H. Xu, et al. 4D-VLA: Spatiotemporal vision-language-action pretraining with cross-scene calibra- tion.arXiv preprint arXiv:2506.22242, 2025

    work page arXiv 2025

  11. [11]

    PaliGemma: A versatile 3B VLM for transfer

    L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, D. Salz, M. Neumann, I. Alabdul- mohsin, M. Tschannen, E. Bugliarello, et al. PaliGemma: A versatile 3b VLM for transfer. arXiv preprint arXiv:2407.07726, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    SmolVLM: Redefining small and efficient multimodal models

    A. Marafioti, O. Zohar, M. Farr ´e, M. Noyan, E. Bakouch, P. Cuenca, C. Zakka, L. B. Allal, A. Lozhkov, N. Tazi, et al. SmolVLM: Redefining small and efficient multimodal models. arXiv preprint arXiv:2504.05299, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al. Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    S. Zhou, A. Vilesov, X. He, Z. Wan, S. Zhang, A. Nagachandra, D. Chang, D. Chen, X. E. Wang, and A. Kadambi. VLM4D: Towards spatiotemporal awareness in vision language mod- els. InProceedings of the IEEE/CVF international conference on computer vision, 2025

    work page 2025

  15. [15]

    Z. He, Y . Wang, C. Zhi, Y . Hu, T.-P. Chen, L. Yin, Z. Chen, T. A. Wu, S. Ouyang, Z. Wang, et al. MemoryArena: Benchmarking agent memory in interdependent multi-session agentic tasks.arXiv preprint arXiv:2602.16313, 2026. 10

    work page arXiv 2026

  16. [16]

    Lauri, D

    M. Lauri, D. Hsu, and J. Pajarinen. Partially observable markov decision processes in robotics: A survey.IEEE Transactions on Robotics, 2022

    work page 2022

  17. [17]

    Y . Wu, S. Liang, C. Zhang, Y . Wang, Y . Zhang, H. Guo, R. Tang, and Y . Liu. From human memory to AI memory: A survey on memory mechanisms in the era of LLMs.arXiv preprint arXiv:2504.15965, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    M. Zhai, Z. Gao, Y . Wu, and Y . Jia. Memory-Centric embodied question answering.arXiv preprint arXiv:2505.13948, 2025

    work page arXiv 2025

  19. [19]

    Sridhar, J

    A. Sridhar, J. Pan, S. Sharma, and C. Finn. MemER: Scaling up memory for robot control via experience retrieval.arXiv preprint arXiv:2510.20328, 2025

    work page arXiv 2025

  20. [20]

    Torne, K

    M. Torne, K. Pertsch, H. Walke, K. Vedder, S. Nair, B. Ichter, A. Z. Ren, H. Wang, J. Tang, K. Stachowicz, et al. MEM: Multi-scale embodied memory for vision language action models. arXiv preprint arXiv:2603.03596, 2026

    work page arXiv 2026

  21. [21]

    T. Chen, Y . Wang, M. Li, Y . Qin, H. Shi, Z. Li, Y . Hu, Y . Zhang, K. Wang, Y . Chen, et al. RM- Bench: Memory-Dependent robotic manipulation benchmark with insights into policy design. arXiv preprint arXiv:2603.01229, 2026

    work page internal anchor Pith review arXiv 2026

  22. [22]

    Y . Wang, Z. Gu, Y . Gao, A. Jiang, Z. Sun, S. Wang, Y . Heng, and H. Sun. HiST-VLA: A hierarchical spatio-temporal vision-language-action model for end-to-end autonomous driving. arXiv preprint arXiv:2602.13329, 2026

    work page arXiv 2026

  23. [23]

    H. Wang, Z. Jing, J. Ao, S. Song, X. Li, G. Huang, and C. Bai. Beyond Short-Horizon: VQ-Memory for robust long-horizon manipulation in non-markovian simulation benchmarks. arXiv preprint arXiv:2603.09513, 2026

    work page arXiv 2026

  24. [24]

    Z. Zeng, F. Ding, H. Yang, and X. Li. HELM: Harness-Enhanced long-horizon memory for vision-language-action manipulation.arXiv preprint arXiv:2604.18791, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    H. Li, F. Shen, D. Chen, L. Yang, X. Wang, J. Shi, Z. Bing, Z. Liu, and A. Knoll. ReMem-VLA: Empowering vision-language-action model with memory via dual-level recurrent queries. arXiv preprint arXiv:2603.12942, 2026

    work page arXiv 2026

  26. [26]

    M. Koo, D. Choi, T. Kim, K. Lee, C. Kim, Y . Seo, and J. Shin. HAMLET: Switch your vision-language-action model into a history-aware policy.arXiv preprint arXiv:2510.00695, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    M. Lin, P. Ding, S. Wang, Z. Zhuang, Y . Liu, X. Tong, W. Song, S. Lyu, S. Huang, and D. Wang. HiF-VLA: Hindsight, Insight and Foresight through motion representation for Vision-Language-Action models.arXiv preprint arXiv:2512.09928, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    J. Liu, Y . Qi, J. Zhang, M. Li, S. Wang, K. Wu, H. Ye, H. Zhang, Z. Chen, F. Zhong, et al. TrackVLA++: Unleashing reasoning and memory capabilities in VLA models for embodied visual tracking.arXiv preprint arXiv:2510.07134, 2025

    work page arXiv 2025

  29. [29]

    Nangue Tasse, M

    G. Nangue Tasse, M. Riemer, B. Rosman, and T. Klinger. Beyond sliding windows: Learning to manage memory in non-markovian environments.arXiv e-prints, 2025

    work page 2025

  30. [30]

    H. Fang, M. Grotz, W. Pumacay, Y . R. Wang, D. Fox, R. Krishna, and J. Duan. SAM2Act: In- tegrating visual foundation model with a memory architecture for robotic manipulation.arXiv preprint arXiv:2501.18564, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Y . Yang, H. Yang, J. Zhou, P. Chen, H. Zhang, Y . Du, and C. Gan. 3D-mem: 3D scene memory for embodied exploration and reasoning. InProceedings of the Computer Vision and Pattern Recognition Conference, 2025. 11

    work page 2025

  32. [32]

    Y . Gao, J. Liu, S. Li, and S. Song. Gated memory policy.arXiv preprint arXiv:2604.18933, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [33]

    Z. Chen, Y . Hu, Z. Fu, Z. Li, J. Huang, Q. Huang, and Y . Wei. INTENT: Invariance and discrimination-aware noise mitigation for robust composed image retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, 2026

    work page 2026

  34. [34]

    Z. Li, Y . Hu, Z. Chen, S. Zhang, Q. Huang, Z. Fu, and Y . Wei. HABIT: Chrono-synergia robust progressive learning framework for composed image retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, 2026

    work page 2026

  35. [35]

    Z. Li, Y . Hu, Z. Chen, Q. Huang, G. Qiu, Z. Fu, and M. Liu. ReTrack: Evidence-driven dual- stream directional anchor calibration network for composed video retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, 2026

    work page 2026

  36. [36]

    A. Zhai, B. Liu, B. Fang, C. Cai, E. Ma, E. Yin, H. Wang, H. Zhou, J. Wang, L. Shi, et al. Igniting VLMs toward the embodied space.arXiv preprint arXiv:2509.11766, 2025

    work page arXiv 2025

  37. [37]

    J. Yang, K. Lin, J. Li, W. Zhang, T. Lin, L. Wu, Z. Su, H. Zhao, Y .-Q. Zhang, L. Chen, et al. RISE: Self-improving robot policy with compositional world model.arXiv preprint arXiv:2602.11075, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  38. [38]

    G. Team, A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, J. Zhu, K. Li, M. Xu, et al. GigaWorld-0: World models as data engine to empower embodied ai.arXiv preprint arXiv:2511.19861, 2025

    work page arXiv 2025

  39. [39]

    G. Team, B. Wang, B. Li, C. Ni, G. Huang, G. Zhao, H. Li, J. Li, J. Lv, J. Liu, et al. GigaBrain- 0.5M∗: a vla that learns from world model-based reinforcement learning.arXiv preprint arXiv:2602.12099, 2026

    work page arXiv 2026

  40. [40]

    A. Ye, Z. Zhang, B. Wang, X. Wang, D. Zhang, and Z. Zhu. VLA-R1: Enhancing reasoning in vision-language-action models.arXiv preprint arXiv:2510.01623, 2025

    work page arXiv 2025

  41. [41]

    H. Li, I. Zhang, R. Ouyang, X. Wang, Z. Zhu, Z. Yang, Z. Zhang, B. Wang, C. Ni, W. Qin, et al. MimicDreamer: Aligning human and robot demonstrations for scalable VLA training. arXiv preprint arXiv:2509.22199, 2025

    work page arXiv 2025

  42. [42]

    Y . Li, L. Zhou, S. Yan, B. Liao, T. Yan, K. Xiong, L. Chen, H. Xie, B. Wang, G. Chen, et al. UniDriveVLA: Unifying understanding, perception, and action planning for autonomous driving.arXiv preprint arXiv:2604.02190, 2026

    work page arXiv 2026

  43. [43]

    J. Wen, Y . Zhu, J. Li, M. Zhu, Z. Tang, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shen, et al. TinyVLA: Towards fast, data-efficient vision-language-action models for robotic manipula- tion.IEEE Robotics and Automation Letters, 2025

    work page 2025

  44. [44]

    DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge

    W. Zhang, H. Liu, Z. Qi, Y . Wang, X. Yu, J. Zhang, R. Dong, J. He, H. Wang, Z. Zhang, et al. DreamVLA: a vision-language-action model dreamed with comprehensive world knowledge. arXiv preprint arXiv:2507.04447, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  45. [45]

    B. Chen, Z. Xu, S. Kirmani, B. Ichter, D. Sadigh, L. Guibas, and F. Xia. SpatialVLM: En- dowing vision-language models with spatial reasoning capabilities. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  46. [46]

    RT-1: Robotics Transformer for Real-World Control at Scale

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, et al. RT-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  47. [47]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  48. [48]

    J. Cen, C. Yu, H. Yuan, Y . Jiang, S. Huang, J. Guo, X. Li, Y . Song, H. Luo, F. Wang, et al. WorldVLA: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    AR-VLA: True Autoregressive Action Expert for Vision-Language-Action Models

    Y . Hu, J.-N. Zaech, N. Nikolov, Y . Yao, S. Dey, G. Albanese, R. Detry, L. Van Gool, and D. Paudel. AR-VLA: True autoregressive action expert for Vision-Language-Action models. arXiv preprint arXiv:2603.10126, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  50. [50]

    Peller-Konrad, R

    F. Peller-Konrad, R. Kartmann, C. R. Dreher, A. Meixner, F. Reister, M. Grotz, and T. As- four. A memory system of a robot cognitive architecture and its implementation in ArmarX. Robotics and Autonomous Systems, 2023

    work page 2023

  51. [51]

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H. Fan, X. Zhang, and G. Huang. MemoryVLA: Perceptual-cognitive memory in vision-language-action models for robotic ma- nipulation.arXiv preprint arXiv:2508.19236, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  52. [52]

    R. Li, W. Guo, Z. Wu, C. Wang, H. Deng, Z. Weng, Y .-P. Tan, and Z. Wang. MAP-VLA: Memory-augmented prompting for vision-language-action model in robotic manipulation. arXiv preprint arXiv:2511.09516, 2025

    work page arXiv 2025

  53. [53]

    M. Neau, Z. Falomir, P. E. Santos, A.-G. Bosser, and C. Buche. GraSP-VLA: Graph-based symbolic action representation for long-horizon planning with VLA policies.arXiv preprint arXiv:2511.04357, 2025

    work page arXiv 2025

  54. [54]

    X. Guo, C. Jiang, H. B. Kim, Y . Sun, Y . Xiao, Y . Han, and J. Yang. Chameleon: Episodic memory for long-horizon robotic manipulation.arXiv preprint arXiv:2603.24576, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  55. [55]

    X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model

    J. Zheng, J. Li, Z. Wang, D. Liu, X. Kang, Y . Feng, Y . Zheng, J. Zou, Y . Chen, J. Zeng, et al. X- VLA: Soft-prompted transformer as scalable cross-embodiment vision-language-action model. arXiv preprint arXiv:2510.10274, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  56. [56]

    Black, A

    K. Black, A. Z. Ren, M. Equi, and S. Levine. Training-time action conditioning for efficient real-time chunking.arXiv preprint arXiv:2512.05964, 2025. 13

    work page arXiv 2025