pith. sign in

arxiv: 2606.12497 · v1 · pith:AWOBCQEFnew · submitted 2026-06-10 · 💻 cs.LG · cs.RO

μVLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models

Egor Cherepanov , Nikita Kachaev , Daniil Zelezetsky , Aydar Bulatov , Artem Pshenitsyn , Yuri Kuratov , Alexey Skrynnik , Aleksandr I. Panov
show 1 more author
Alexey K. Kovalev
This is my paper

Pith reviewed 2026-06-27 10:48 UTC · model grok-4.3

classification 💻 cs.LG cs.RO
keywords vision-language-actionrecurrent memorypartial observabilitymanipulation taskstransformermemory tokensVLA models
0
0 comments X

The pith

Adding a small set of learnable memory tokens to a pretrained VLA transformer, updated only through self-attention and trained with TBPTT, raises average success from 0.42 to 0.84 on partially observable manipulation tasks.

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

The paper isolates recurrence as the sole variable by inserting learnable memory tokens into an existing VLA backbone; these tokens persist across time steps and are updated inside the transformer's self-attention layers. Training uses truncated backpropagation through time with no auxiliary losses or extra modules. On a benchmark containing both training and held-out tasks that share the same memory demands, the recurrent versions reach substantially higher success rates than the memoryless baseline, while tasks with mismatched memory demands show no improvement and fully observable tasks show no regression. A reader would care because the design cleanly separates the contribution of recurrence from retrieval, compression, or hierarchical memory, revealing the regime in which minimal in-backbone recurrence is already sufficient.

Core claim

By augmenting a transformer-based VLA with a small set of learnable memory tokens carried across timesteps and updated through self-attention, trained end-to-end with truncated backpropagation through time and no auxiliary losses, the resulting models achieve average success rates of 0.84 on five training tasks and 0.23 on held-out tasks that share the identical memory structure, compared with 0.42 and 0.07 for the memoryless baseline; performance stays near baseline on tasks requiring a different memory structure and reaches 96.2 percent average success on a fully observable benchmark.

What carries the argument

A small set of learnable memory tokens carried across timesteps and updated through self-attention inside the transformer, with the only varying factors being memory width, TBPTT length, and the choice between cross-step gradients or detached EMA update.

If this is right

  • On tasks sharing the same memory demand structure, recurrent models reach substantially higher success rates than memoryless baselines.
  • On tasks whose memory demands differ, recurrent models perform near the memoryless baseline.
  • The strongest recurrent variant reaches 96.2 percent average success on fully observable benchmarks with no regression.
  • Recurrence can be isolated from retrieval, compression, auxiliary objectives, or hierarchical memory while still producing measurable gains.

Where Pith is reading between the lines

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

  • For many manipulation tasks whose partial observability fits a simple temporal pattern, minimal recurrence may be sufficient and more complex memory architectures may be unnecessary.
  • The gap between training and held-out tasks with matched versus mismatched memory structure suggests that generalization depends on correctly identifying the required memory pattern in advance.
  • Extending the same token-based recurrence to longer horizons or different sensor modalities could test the boundaries of the regime identified here.
  • The same isolation approach could be applied to other sequential decision domains where partial observability is the limiting factor.

Load-bearing premise

The measured performance gains are produced by the recurrence mechanism itself rather than by differences in optimization dynamics, task selection, or the labeling of which tasks require memory.

What would settle it

An experiment that matches training procedure exactly and still finds no success-rate difference between recurrent and memoryless models on the same partially observable tasks would falsify the claim that recurrence drives the gains.

Figures

Figures reproduced from arXiv: 2606.12497 by Aleksandr I. Panov, Alexey K. Kovalev, Alexey Skrynnik, Artem Pshenitsyn, Aydar Bulatov, Daniil Zelezetsky, Egor Cherepanov, Nikita Kachaev, Yuri Kuratov.

Figure 1
Figure 1. Figure 1: µVLA: recurrent memory inside a pretrained VLA backbone. At each timestep t, the model consumes visual observations, proprioception, language, and memory tokens from the previous step, then predicts actions and updated memory tokens passed recurrently to t + 1. Abstract Vision-language-action (VLA) models predict chunks of future actions from the current observation, an assumption that fails under partial … view at source ↗
Figure 2
Figure 2. Figure 2: TBPTT training of µVLA. The pretrained VLA is unrolled over a window of K steps. Memory tokens are initialized from Minit at t0 and recurrently passed between successive steps. Gradients flow through the unrolled pathway and are detached at the window boundary. 1 Introduction Vision-language-action (VLA) models cast manipulation as next-token prediction over multimodal sequences [79, 27, 28, 3, 24, 19, 2, … view at source ↗
Figure 3
Figure 3. Figure 3: Attention mask with the memory-action guard. Memory tokens attend only to observations, proprioception, language, and previous memory state, but cannot read ac￾tion tokens. This prevents the recurrent state from trivially copying demonstrated actions and encourages encoding of task-relevant observations instead. Incorporating history in VLA mod￾els. Prior work incorporates past in￾formation in three main w… view at source ↗
Figure 4
Figure 4. Figure 4: Cosine distance Mt vs. Mt−1 on RememberColor5 (K=2, m=64). Pale blue lines: individ￾ual memory tokens; dark grey: mean over all 64. The episode contains two phase transitions: the cue cube is visible until t=4; at t=5 it disappears and the agent observes an empty table until t=9, while at t=10 the candidate cubes appear and the task switches from memory maintenance to retrieval and action execution. Colors… view at source ↗
Figure 5
Figure 5. Figure 5: Memory intervention on RememberColor5. SR at 100 episodes un￾der baseline, noise, and freeze_first. Composing per-layer attention with rollout (mean over heads, residual factor 0.5) on three query groups (action→vision, memory→vision, vision→vision) yields overlays (Appendix K) in which action attention tracks the gripper and target object, while memory at￾tention concentrates on the cue region before mask… view at source ↗
Figure 6
Figure 6. Figure 6: Memory→vision attention rollout on RememberColor5 (K=2, m=64). The cue panel is visible early in the episode and then masked. Memory attention concentrates on the cue grid while it is visible and becomes diffuse after occlusion, matching the phase transition in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Round-robin episodic dataloader. Each batch slot b ∈ {0, . . . , B−1} is assigned an independent stream that walks one episode step by step before sampling a new episode. Steps are emitted in fixed round-robin order so that slot b in two consecutive batches always belongs to the same stream. The recurrent state for slot b accumulated at batch τ is therefore the correct input at batch τ+1. An is_first flag … view at source ↗
Figure 8
Figure 8. Figure 8: aggregates the phase-length and color-swap generalization curves for RememberColor5 [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Memory dynamics on RememberColor5 (K=2, m=64). Companion to [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Memory dynamics on TakeItBack (K=2, m=64). The cosine trace shows a two-event write pattern: the canonical t=1 spike (∆ cos≈0.18) and a second spike at t≈18 when the policy switches to the return phase. The mean L2 norm rises from ≈ 40 to ≈ 75, with a sharp jump aligned with the same phase transition. M MIKASA-Robo Training Environment Descriptions RememberColor5-VLA-v0 At the beginning of each episode, o… view at source ↗
Figure 11
Figure 11. Figure 11: Memory dynamics on ShellGamePush (K=2, m=64). A single early writing window: a large t=1 cosine spike (∆ cos≈0.27) followed by a small secondary bump at t=4 (≈ 0.16) and a quiet ≈ 0.10 plateau thereafter. This is consistent with the cue being captured in the first frames and then carried largely unchanged through the push. O Extended Related Work Vision-language-action models. RT-1 [4] introduced a scalab… view at source ↗
Figure 12
Figure 12. Figure 12: Memory dynamics on InterceptMedium (K=2, m=64). Continuous re-grounding rather than a one-shot write: the t=1 spike (∆ cos≈0.12) is followed by a sustained mid-episode plateau at ≈ 0.06–0.08. The mean L2 norm shows a pronounced bump at t=15–18 (≈ 75 vs. ≈ 45 baseline) aligned with the catch event. working memory with a transformer-based episodic compressor that recursively summarizes older observations in… view at source ↗
Figure 13
Figure 13. Figure 13: Memory dynamics on RememberShapeAndColor3x3 (K=2, m=64). Two-phase write: a t=1 spike (∆ cos ≈ 0.28) for initial encoding, then a secondary bump at t=5–7 (≈ 0.13) when the cue is fully revealed and committed. After t≈10 the cosine settles to ≈ 0.05 and the L2 norm plateaus at ≈ 60. Horizons [26] jointly processes several action-chunk horizons, VLA Knows Its Limits [63] adapts the execution horizon at test… view at source ↗
Figure 14
Figure 14. Figure 14: Attention rollouts on RememberColor5 (K=2, m=64). Per-step attention rollout over the 32 backbone layers, mean over heads, normalised per step. Row (a) is the raw episode (top and wrist cameras stacked). Rows (b)–(d) overlay the rolled-out attention from three query groups onto the same frames: the first action token, the memory tokens, and the vision tokens themselves. The action query tracks the gripper… view at source ↗
Figure 15
Figure 15. Figure 15: Attention rollouts on RememberShapeAndColor3x3 (K=2, m=64). The cue panel is a 3 × 3 shape-by-colour grid that is fully revealed early and then masked. The memory→vision row places its peaks on the grid cells while they are visible and broadens once the panel is occluded, matching the two-phase write event in [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Attention rollouts on ShellGamePush (K=2, m=64). The episode has a single-write cue dynamic: the ball is visible only in the first frames before the cup occludes it, and the cosine trace in [PITH_FULL_IMAGE:figures/full_fig_p031_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Attention rollouts on TakeItBack (K=2, m=64). The task has two phases (move out, return), and the cosine trace in [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Attention rollouts on InterceptMedium (K=2, m=64). The task requires continuous re-grounding rather than a one-shot write, and the cosine trace in [PITH_FULL_IMAGE:figures/full_fig_p032_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Memory intervention on the five MIKASA-Robo training tasks. The noise bar is below the baseline in every cell with a non-trivial baseline, confirming that the recurrent channel is functionally read at inference. freeze_first retains part of the SR on cue-recall tasks where the cue is localised in the first frames (RC5, TakeItBack with K=2), and collapses on dynamics-grounded tasks (InterceptMedium). Compa… view at source ↗
Figure 20
Figure 20. Figure 20: Chunk-length sweep on the five MIKASA-Robo training tasks. For each carrier (K=8, EMA, K=2) we report SR at chunk lengths {1, 2, 4, 8, 16} with the memory channel active (solid) and zeroed at inference (dashed). chunk=1 is the receding-horizon regime used at training. The with-memory minus no-memory gap is largest at chunk=1 on cue-recall tasks (RC5, RemSAC3x3, TakeItBack), while long-chunk inference part… view at source ↗
read the original abstract

Vision-language-action (VLA) models predict chunks of future actions from the current observation, an assumption that fails under partial observability, where decisions depend on information no longer visible. Existing memory-augmented VLAs simultaneously introduce recurrence, retrieval, compression modules, auxiliary objectives, hierarchical memory, or task-specific architectural changes, so the contribution of recurrence itself remains entangled with surrounding machinery. We present a controlled isolation study of recurrence in a strong pretrained VLA backbone. Our formulation augments the transformer with a small set of learnable memory tokens carried across timesteps and updated through self-attention, trained end to end with truncated backpropagation through time, with no auxiliary losses and no architectural changes. We instantiate this as $\mu$VLA, a family of OpenVLA-OFT variants parameterized by memory width m, TBPTT length K, and the memory update rule (cross-step gradients or a detached EMA), so that recurrence is the only varying factor. On MIKASA-Robo, $\mu$VLA improves average success rate on five training tasks from 0.42 to 0.84 at the strongest setting and reaches 0.23 on held-out tasks with the same memory structure versus 0.07 for the memoryless baseline. On tasks requiring different memory structure, performance remains near baseline. On LIBERO, the strongest recurrent variant achieves 96.2% average success, indicating no regression under full observability. We interpret these results as a calibration of the capability envelope of minimal in-backbone recurrence, identifying the regime in which it is sufficient and the regime where additional memory structure is required. Demos and videos can be found in https://avanturist322.github.io/mu-vla/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper claims to isolate the contribution of recurrence in VLA models for partially observable manipulation by augmenting a pretrained backbone (OpenVLA-OFT) with a small set of learnable memory tokens that are carried across timesteps, updated via self-attention, and trained end-to-end with TBPTT (no auxiliary losses or other architectural changes). Recurrence is the sole varying factor via parameters m (memory width), K (TBPTT length), and update rule (cross-step gradients or detached EMA). On MIKASA-Robo this yields average success rate gains from 0.42 to 0.84 on five training tasks at the strongest setting and from 0.07 to 0.23 on held-out tasks sharing the same memory structure (versus memoryless baseline), with performance remaining near baseline on tasks requiring different memory structure; on LIBERO the strongest variant reaches 96.2% average success with no regression under full observability. The results are interpreted as calibrating the regime in which minimal in-backbone recurrence suffices.

Significance. If the isolation of recurrence holds, the work supplies a useful empirical calibration of the capability envelope of minimal recurrence inside VLA transformers for POMDPs, distinguishing regimes where the mechanism is sufficient from those requiring additional memory structure. The explicit parameterization by m, K, and update rule, together with the attempt at a controlled study on a strong pretrained backbone, is a methodological strength that could serve as a reproducible baseline for future memory-augmented VLA research.

major comments (1)
  1. [MIKASA-Robo experiments and held-out task evaluation] The central claim that performance deltas (0.42→0.84 on training tasks, 0.07→0.23 on held-out) are attributable to recurrence rather than extra capacity, attention changes, or task selection requires that the assignment of tasks to 'same memory structure' versus 'different memory structure' be based on a pre-specified, observable criterion (e.g., a formal Markovian/non-Markovian state definition or other independent property) defined before seeing results. The manuscript provides no such independent criterion in the experimental section, leaving the isolation claim under-determined.
minor comments (2)
  1. [Results] The abstract and results statements report point estimates without error bars, number of seeds, or statistical tests; these should be added to all tables and figures reporting success rates.
  2. [Experiments] Full baseline details, ablation tables over m and K, and exact task definitions for the memory-structure split should be expanded for reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive and detailed review. We address the single major comment below and will revise the manuscript to strengthen the experimental claims.

read point-by-point responses

  1. Referee: [MIKASA-Robo experiments and held-out task evaluation] The central claim that performance deltas (0.42→0.84 on training tasks, 0.07→0.23 on held-out) are attributable to recurrence rather than extra capacity, attention changes, or task selection requires that the assignment of tasks to 'same memory structure' versus 'different memory structure' be based on a pre-specified, observable criterion (e.g., a formal Markovian/non-Markovian state definition or other independent property) defined before seeing results. The manuscript provides no such independent criterion in the experimental section, leaving the isolation claim under-determined.

    Authors: We agree that the isolation claim would be strengthened by an explicit, pre-specified criterion for task categorization that is independent of observed model performance. The current manuscript does not provide such a criterion in the experimental section. In the revised manuscript we will add a dedicated subsection (prior to the results) that defines 'same memory structure' versus 'different' based on an observable, benchmark-defined property: whether the task requires the agent to maintain information about object identities or positions that become fully occluded after the initial observation (as specified in the MIKASA-Robo task definitions). This definition will be stated before any training or evaluation results are presented, ensuring the categorization is not post-hoc. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparisons only

full rationale

The paper reports experimental success rates (0.42→0.84 on training tasks, 0.23 vs 0.07 on held-out) from controlled variants of memory-token recurrence in a pretrained VLA backbone. No equations, fitted parameters presented as predictions, or derivation steps exist that reduce any claimed result to its inputs by construction. Task descriptions and memory-structure labels are given as experimental conditions; results are framed as direct measurements rather than self-referential outputs. Any self-citations are incidental and not load-bearing for a mathematical claim.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

Only the abstract is available, so the ledger is limited to elements explicitly named there; memory tokens are introduced as a new mechanism whose effectiveness is demonstrated empirically rather than derived.

free parameters (2)
  • memory width m
    Number of learnable memory tokens; treated as a tunable hyperparameter that defines model variants.
  • TBPTT length K
    Truncated backpropagation through time horizon; chosen per variant.
axioms (1)
  • domain assumption Self-attention over memory tokens can maintain and update task-relevant state across timesteps without auxiliary losses.
    Invoked when the paper states that memory tokens are updated through self-attention and trained end-to-end.
invented entities (1)
  • learnable memory tokens no independent evidence
    purpose: Carry information across timesteps inside the transformer backbone for partial observability.
    New component added to the pretrained VLA; no independent evidence outside the reported experiments is provided in the abstract.

pith-pipeline@v0.9.1-grok · 5899 in / 1606 out tokens · 31679 ms · 2026-06-27T10:48:50.977293+00:00 · methodology

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

80 extracted references · 7 canonical work pages

  1. [1]

    Recursive belief vision language action models.arXiv preprint arXiv:2602.20659, 2026

    Vaidehi Bagaria, Bijo Sebastian, and Nirav Kumar Patel. Recursive belief vision language action models.arXiv preprint arXiv:2602.20659, 2026

    arXiv 2026

  2. [2]

    Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    and Kragic, D

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Robert Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, Laura Smith, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Z...

    work page doi:10.15607/rss.2025.xxi.010 2025

  4. [4]

    Brohan, N

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Mall...

    work page doi:10.15607/rss.2023.xix.025 2023

  5. [5]

    Recurrent memory transformer.Advances in Neural Information Processing Systems, 35:11079–11091, 2022

    Aydar Bulatov, Yury Kuratov, and Mikhail Burtsev. Recurrent memory transformer.Advances in Neural Information Processing Systems, 35:11079–11091, 2022

    2022

  6. [6]

    History-aware visuomotor policy learning via point tracking.arXiv preprint arXiv:2509.17141, 2025

    Jingjing Chen, Hongjie Fang, Chenxi Wang, Shiquan Wang, and Cewu Lu. History-aware visuomotor policy learning via point tracking.arXiv preprint arXiv:2509.17141, 2025

    arXiv 2025

  7. [7]

    Rmbench: Memory-dependent robotic manipulation benchmark with insights into policy design.arXiv preprint arXiv:2603.01229, 2026

    Tianxing Chen, Yuran Wang, Mingleyang Li, Yan Qin, Hao Shi, Zixuan Li, Yifan Hu, Yingsheng Zhang, Kaixuan Wang, Yue Chen, et al. Rmbench: Memory-dependent robotic manipulation benchmark with insights into policy design.arXiv preprint arXiv:2603.01229, 2026

    arXiv 2026

  8. [8]

    Longbench: Evaluating robotic manipulation policies on real-world long-horizon tasks.arXiv preprint arXiv:2604.16788, 2026

    Xueyao Chen, Jingkai Jia, Tong Yang, Yibo Fu, Wei Li, and Wenqiang Zhang. Longbench: Evaluating robotic manipulation policies on real-world long-horizon tasks.arXiv preprint arXiv:2604.16788, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Robohiman: A hierarchical evaluation paradigm for compositional generalization in long-horizon manipulation.arXiv preprint arXiv:2510.13149, 2025

    Yangtao Chen, Zixuan Chen, Nga Teng Chan, Junting Chen, Junhui Yin, Jieqi Shi, Yang Gao, Yong-Lu Li, and Jing Huo. Robohiman: A hierarchical evaluation paradigm for compositional generalization in long-horizon manipulation.arXiv preprint arXiv:2510.13149, 2025

    arXiv 2025

  10. [10]

    Non-markovian long-horizon robot manipulation via keyframe chaining.arXiv preprint arXiv:2603.01465, 2026

    Yipeng Chen, Wentao Tan, Lei Zhu, Fengling Li, Jingjing Li, Guoli Yang, and Heng Tao Shen. Non-markovian long-horizon robot manipulation via keyframe chaining.arXiv preprint arXiv:2603.01465, 2026

    arXiv 2026

  11. [11]

    Memory, bench- mark & robots: A benchmark for solving complex tasks with reinforcement learning

    Egor Cherepanov, Nikita Kachaev, Alexey Kovalev, and Aleksandr Panov. Memory, bench- mark & robots: A benchmark for solving complex tasks with reinforcement learning. In The Fourteenth International Conference on Learning Representations, 2026. URL https: //openreview.net/forum?id=9cLPurIZMj

    2026

  12. [12]

    ELMUR: External layer memory with update/rewrite for long-horizon RL problems

    Egor Cherepanov, Alexey Kovalev, and Aleksandr Panov. ELMUR: External layer memory with update/rewrite for long-horizon RL problems. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=bm3rbtEMFj

    2026

  13. [13]

    Recurrent action transformer with memory

    Egor Cherepanov, Aleksei Staroverov, Alexey Kovalev, and Aleksandr Panov. Recurrent action transformer with memory. InThe Fourteenth International Conference on Learning Representations, 2026. URLhttps://openreview.net/forum?id=kByN4v0M3e. 10

    2026

  14. [14]

    Diffusion policy: Visuomotor policy learning via action diffusion

    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. The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    2025

  15. [15]

    Rethinking progression of memory state in robotic manipulation: An object-centric perspective

    Nhat Chung, Taisei Hanyu, Toan Nguyen, Huy Le, Frederick Bumgarner, Duy Minh Ho Nguyen, Khoa V o, Kashu Yamazaki, Chase Rainwater, Tung Kieu, et al. Rethinking progression of memory state in robotic manipulation: An object-centric perspective. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 3407–3415, 2026

    2026

  16. [16]

    Robomme: Benchmarking and understanding memory for robotic generalist policies.arXiv preprint arXiv:2603.04639, 2026

    Yinpei Dai, Hongze Fu, Jayjun Lee, Yuejiang Liu, Haoran Zhang, Jianing Yang, Chelsea Finn, Nima Fazeli, and Joyce Chai. Robomme: Benchmarking and understanding memory for robotic generalist policies.arXiv preprint arXiv:2603.04639, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Long-vla: Unleashing long-horizon capability of vision language action model for robot manipulation.arXiv preprint arXiv:2508.19958, 2025

    Yiguo Fan, Pengxiang Ding, Shuanghao Bai, Xinyang Tong, Yuyang Zhu, Hongchao Lu, Fengqi Dai, Wei Zhao, Yang Liu, Siteng Huang, et al. Long-vla: Unleashing long-horizon capability of vision language action model for robot manipulation.arXiv preprint arXiv:2508.19958, 2025

    arXiv 2025

  18. [18]

    Gated memory policy.arXiv preprint arXiv:2604.18933, 2026

    Yihuai Gao, Jinyun Liu, Shuang Li, and Shuran Song. Gated memory policy.arXiv preprint arXiv:2604.18933, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Octo: An Open-Source Generalist Robot Policy

    Dibya Ghosh, Homer Rich Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Quan Vuong, Ted Xiao, Pannag R Sanketi, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An Open-Source Generalist Robot Policy. InProceedings of Robotics: Science and Systems, ...

    work page doi:10.15607/rss.2024.xx.090 2024

  20. [20]

    Chameleon: Episodic memory for long-horizon robotic manipulation.arXiv preprint arXiv:2603.24576, 2026

    Xinying Guo, Chenxi Jiang, Hyun Bin Kim, Ying Sun, Yang Xiao, Yuhang Han, and Jianfei Yang. Chameleon: Episodic memory for long-horizon robotic manipulation.arXiv preprint arXiv:2603.24576, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    Robocerebra: A large-scale benchmark for long-horizon robotic manipulation evaluation.arXiv preprint arXiv:2506.06677, 2025

    Songhao Han, Boxiang Qiu, Yue Liao, Siyuan Huang, Chen Gao, Shuicheng Yan, and Si Liu. Robocerebra: A large-scale benchmark for long-horizon robotic manipulation evaluation.arXiv preprint arXiv:2506.06677, 2025

    arXiv 2025

  22. [22]

    Notes-to- self: Scratchpad augmented vlas for memory dependent manipulation tasks.arXiv preprint arXiv:2602.21013, 2026

    Sanjay Haresh, Daniel Dijkman, Apratim Bhattacharyya, and Roland Memisevic. Notes-to- self: Scratchpad augmented vlas for memory dependent manipulation tasks.arXiv preprint arXiv:2602.21013, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=nZeVKeeFYf9

    2022

  24. [24]

    π0.5: a vision- language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. π0.5: a vision- language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Contextvla: Vision-language-action model with amortized multi-frame context.arXiv preprint arXiv:2510.04246, 2025

    Huiwon Jang, Sihyun Yu, Heeseung Kwon, Hojin Jeon, Younggyo Seo, and Jinwoo Shin. Contextvla: Vision-language-action model with amortized multi-frame context.arXiv preprint arXiv:2510.04246, 2025

    arXiv 2025

  26. [26]

    Mixture of horizons in action chunking.arXiv preprint arXiv:2511.19433, 2025

    Dong Jing, Gang Wang, Jiaqi Liu, Weiliang Tang, Zelong Sun, Yunchao Yao, Zhenyu Wei, Yunhui Liu, Zhiwu Lu, and Mingyu Ding. Mixture of horizons in action chunking.arXiv preprint arXiv:2511.19433, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Open- VLA: An open-source vision-language-action model

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Open- VLA: An open-source vision-language-action model. In8th Annual Conference on Robot Learn...

    2024

  28. [28]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. InProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025. doi: 10.15607/RSS.2025.XXI.017

    work page doi:10.15607/rss.2025.xxi.017 2025

  29. [29]

    Hamlet: Switch your vision-language-action model into a history-aware policy.arXiv preprint arXiv:2510.00695, 2025

    Myungkyu Koo, Daewon Choi, Taeyoung Kim, Kyungmin Lee, Changyeon Kim, Younggyo Seo, and Jinwoo Shin. Hamlet: Switch your vision-language-action model into a history-aware policy.arXiv preprint arXiv:2510.00695, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Robomemory: A brain-inspired multi-memory agentic framework for interactive environmental learning in physical embodied systems.arXiv preprint arXiv:2508.01415, 2025

    Mingcong Lei, Honghao Cai, Yuyuan Yang, Yimou Wu, Jinke Ren, Zezhou Cui, Liangchen Tan, Junkun Hong, Gehan Hu, Shuangyu Zhu, et al. Robomemory: A brain-inspired multi-memory agentic framework for interactive environmental learning in physical embodied systems.arXiv preprint arXiv:2508.01415, 2025

    arXiv 2025

  31. [31]

    Vpwem: Non-markovian visuomotor policy with working and episodic memory.arXiv preprint arXiv:2603.04910, 2026

    Yuheng Lei, Zhixuan Liang, Hongyuan Zhang, and Ping Luo. Vpwem: Non-markovian visuomotor policy with working and episodic memory.arXiv preprint arXiv:2603.04910, 2026

    arXiv 2026

  32. [32]

    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

  33. [33]

    Cronusvla: Transferring latent motion across time for multi-frame prediction in manipulation.arXiv preprint arXiv:2506.19816, 2025

    Hao Li, Shuai Yang, Yilun Chen, Yang Tian, Xiaoda Yang, Xinyi Chen, Hanqing Wang, Tai Wang, Feng Zhao, Dahua Lin, et al. Cronusvla: Transferring latent motion across time for multi-frame prediction in manipulation.arXiv preprint arXiv:2506.19816, 2025

    arXiv 2025

  34. [34]

    Physmem: Scaling test-time physical memory for robot manipulation.arXiv preprint arXiv:2602.20323, 2026

    Haoyang Li, Yang You, Hao Su, and Leonidas Guibas. Physmem: Scaling test-time physical memory for robot manipulation.arXiv preprint arXiv:2602.20323, 2026

    Pith/arXiv arXiv 2026

  35. [35]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation.arXiv preprint arXiv:2411.19650, 2024

    Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation.arXiv preprint arXiv:2411.19650, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Map-vla: Memory-augmented prompting for vision-language-action model in robotic manipulation.arXiv preprint arXiv:2511.09516, 2025

    Runhao Li, Wenkai Guo, Zhenyu Wu, Changyuan Wang, Haoyuan Deng, Zhenyu Weng, Yap- Peng Tan, and Ziwei Wang. Map-vla: Memory-augmented prompting for vision-language-action model in robotic manipulation.arXiv preprint arXiv:2511.09516, 2025

    arXiv 2025

  37. [37]

    Global prior meets local consistency: Dual-memory augmented vision- language-action model for efficient robotic manipulation.arXiv preprint arXiv:2602.20200, 2026

    Zaijing Li, Bing Hu, Rui Shao, Gongwei Chen, Dongmei Jiang, Pengwei Xie, Jianye Hao, and Liqiang Nie. Global prior meets local consistency: Dual-memory augmented vision- language-action model for efficient robotic manipulation.arXiv preprint arXiv:2602.20200, 2026

    Pith/arXiv arXiv 2026

  38. [38]

    Echovla: Robotic vision-language-action model with synergistic declarative memory for mobile manipulation.arXiv preprint arXiv:2511.18112, 2025

    Min Lin, Xiwen Liang, Bingqian Lin, Liu Jingzhi, Zijian Jiao, Kehan Li, Yuhan Ma, Yuecheng Liu, Shen Zhao, Yuzheng Zhuang, et al. Echovla: Robotic vision-language-action model with synergistic declarative memory for mobile manipulation.arXiv preprint arXiv:2511.18112, 2025

    arXiv 2025

  39. [39]

    Hif-vla: Hindsight, insight and foresight through motion representation for vision-language-action models.arXiv preprint arXiv:2512.09928, 2025

    Minghui Lin, Pengxiang Ding, Shu Wang, Zifeng Zhuang, Yang Liu, Xinyang Tong, Wenxuan Song, Shangke Lyu, Siteng Huang, and Donglin Wang. Hif-vla: Hindsight, insight and foresight through motion representation for vision-language-action models.arXiv preprint arXiv:2512.09928, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    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

    2023

  41. [41]

    Long-horizon manipulation via trace-conditioned vla planning.arXiv preprint arXiv:2604.21924, 2026

    Isabella Liu, An-Chieh Cheng, Rui Yan, Geng Chen, Ri-Zhao Qiu, Xueyan Zou, Sha Yi, Hongxu Yin, Xiaolong Wang, and Sifei Liu. Long-horizon manipulation via trace-conditioned vla planning.arXiv preprint arXiv:2604.21924, 2026

    Pith/arXiv arXiv 2026

  42. [42]

    Evovla: Self-evolving vision-language- action model.arXiv preprint arXiv:2511.16166, 2025

    Zeting Liu, Zida Yang, Zeyu Zhang, and Hao Tang. Evovla: Self-evolving vision-language- action model.arXiv preprint arXiv:2511.16166, 2025. 12

    arXiv 2025

  43. [43]

    Goal2skill: Long-horizon manipulation with adaptive planning and reflection.arXiv preprint arXiv:2604.13942, 2026

    Zhen Liu, Xinyu Ning, Zhe Hu, Xinxin Xie, Weize Li, Zhipeng Tang, Chongyu Wang, Zejun Yang, Hanlin Wang, Yitong Liu, et al. Goal2skill: Long-horizon manipulation with adaptive planning and reflection.arXiv preprint arXiv:2604.13942, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    F1: A vision-language-action model bridging understanding and generation to actions.arXiv preprint arXiv:2509.06951, 2025

    Qi Lv, Weijie Kong, Hao Li, Jia Zeng, Zherui Qiu, Delin Qu, Haoming Song, Qizhi Chen, Xiang Deng, and Jiangmiao Pang. F1: A vision-language-action model bridging understanding and generation to actions.arXiv preprint arXiv:2509.06951, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Bpp: Long-context robot imitation learning by focusing on key history frames.arXiv preprint arXiv:2602.15010, 2026

    Max Sobol Mark, Jacky Liang, Maria Attarian, Chuyuan Fu, Debidatta Dwibedi, Dhruv Shah, and Aviral Kumar. Bpp: Long-context robot imitation learning by focusing on key history frames.arXiv preprint arXiv:2602.15010, 2026

    arXiv 2026

  46. [46]

    Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks.IEEE Robotics and Automation Letters, 7(3):7327–7334, 2022

    2022

  47. [47]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick L...

    2024

  48. [48]

    Open x- embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0

    Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x- embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

    2024

  49. [49]

    Spatial traces: Enhancing vla models with spatial-temporal understanding.Optical Memory and Neural Networks, 34(Suppl 1):S72–S82, 2025

    Maxim A Patratskiy, Alexey K Kovalev, and Aleksandr I Panov. Spatial traces: Enhancing vla models with spatial-temporal understanding.Optical Memory and Neural Networks, 34(Suppl 1):S72–S82, 2025

    2025

  50. [50]

    FAST: Efficient Action Tokenization for Vision-Language- Action Models

    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. InProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June

  51. [51]

    doi: 10.15607/RSS.2025.XXI.012

    work page doi:10.15607/rss.2025.xxi.012 2025

  52. [52]

    Efficient long-horizon vision-language-action models via static-dynamic disentanglement.arXiv preprint arXiv:2602.03983, 2026

    Weikang Qiu, Tinglin Huang, and Rex Ying. Efficient long-horizon vision-language-action models via static-dynamic disentanglement.arXiv preprint arXiv:2602.03983, 2026

    Pith/arXiv arXiv 2026

  53. [53]

    SpatialVLA: Exploring Spatial Representa- tions for Visual-Language-Action Models

    Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Jiayuan Gu, Zhigang Wang, Yan Ding, Bin Zhao, Dong Wang, and Xuelong Li. SpatialVLA: Exploring Spatial Representa- tions for Visual-Language-Action Models. InProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025. doi: 10.15607/RSS.2025.XXI.011

    work page doi:10.15607/rss.2025.xxi.011 2025

  54. [54]

    Scaling short-term memory of visuo- motor policies for long-horizon tasks, 2026

    Rutav Shah, Rajat Kumar Jenamani, Xiaohan Zhang, Lingfeng Sun, Roberto Martín-Martín, Yuke Zhu, Deva Ramanan, and Karl Schmeckpeper. Scaling short-term memory of visuo- motor policies for long-horizon tasks, 2026. URL https://openreview.net/forum?id= 5SMNtmJFGa

    2026

  55. [55]

    Memoryvla: Perceptual-cognitive memory in vision- language-action models for robotic manipulation.arXiv preprint arXiv:2508.19236, 2025

    Hao Shi, Bin Xie, Yingfei Liu, Lin Sun, Fengrong Liu, Tiancai Wang, Erjin Zhou, Haoqiang Fan, Xiangyu Zhang, and Gao Huang. Memoryvla: Perceptual-cognitive memory in vision- language-action models for robotic manipulation.arXiv preprint arXiv:2508.19236, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    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

  57. [57]

    Memer: Scaling up memory for robot control via experience retrieval.arXiv preprint arXiv:2510.20328, 2025

    Ajay Sridhar, Jennifer Pan, Satvik Sharma, and Chelsea Finn. Memer: Scaling up memory for robot control via experience retrieval.arXiv preprint arXiv:2510.20328, 2025. 13

    arXiv 2025

  58. [58]

    Tempofit: Plug-and-play layer-wise tem- poral kv memory for long-horizon vision-language-action manipulation.arXiv preprint arXiv:2603.07647, 2026

    Jun Sun, Boyu Yang, Jiahao Zhang, Ning Ma, Chencheng Wu, Siqing Zhang, Yiou Huang, Qiufeng Wang, Shan Liang, and Yaran Chen. Tempofit: Plug-and-play layer-wise tem- poral kv memory for long-horizon vision-language-action manipulation.arXiv preprint arXiv:2603.07647, 2026

    arXiv 2026

  59. [59]

    Action-sketcher: From reasoning to ac- tion via visual sketches for long-horizon robotic manipulation.arXiv preprint arXiv:2601.01618, 2026

    Huajie Tan, Peterson Co, Yijie Xu, Shanyu Rong, Yuheng Ji, Cheng Chi, Xiansheng Chen, Qiongyu Zhang, Zhongxia Zhao, Pengwei Wang, et al. Action-sketcher: From reasoning to ac- tion via visual sketches for long-horizon robotic manipulation.arXiv preprint arXiv:2601.01618, 2026

    arXiv 2026

  60. [60]

    Mem: Multi-scale embodied memory for vision language action models.arXiv preprint arXiv:2603.03596, 2026

    Marcel Torne, Karl Pertsch, Homer Walke, Kyle Vedder, Suraj Nair, Brian Ichter, Allen Z Ren, Haohuan Wang, Jiaming Tang, Kyle Stachowicz, et al. Mem: Multi-scale embodied memory for vision language action models.arXiv preprint arXiv:2603.03596, 2026

    arXiv 2026

  61. [61]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  62. [62]

    Recurrent-depth vla: Implicit test-time compute scaling of vision-language-action models via latent iterative reasoning.arXiv preprint arXiv:2602.07845, 2026

    Yalcin Tur, Jalal Naghiyev, Haoquan Fang, Wei-Chuan Tsai, Jiafei Duan, Dieter Fox, and Ranjay Krishna. Recurrent-depth vla: Implicit test-time compute scaling of vision-language-action models via latent iterative reasoning.arXiv preprint arXiv:2602.07845, 2026

    arXiv 2026

  63. [63]

    Codegraphvlp: Code-as-planner meets semantic-graph state for non-markovian vision-language-action models.arXiv preprint arXiv:2604.22238, 2026

    Khoa V o, Sieu Tran, Taisei Hanyu, Yuki Ikebe, Duy Nguyen, Bui Duy Quoc Nghi, Minh Vu, Anthony Gunderman, Chase Rainwater, Anh Nguyen, et al. Codegraphvlp: Code-as-planner meets semantic-graph state for non-markovian vision-language-action models.arXiv preprint arXiv:2604.22238, 2026

    Pith/arXiv arXiv 2026

  64. [64]

    Vla knows its limits.arXiv preprint arXiv:2602.21445, 2026

    Haoxuan Wang, Gengyu Zhang, Yan Yan, Ramana Rao Kompella, and Gaowen Liu. Vla knows its limits.arXiv preprint arXiv:2602.21445, 2026

    Pith/arXiv arXiv 2026

  65. [65]

    Beyond short-horizon: Vq-memory for robust long-horizon manipulation in non-markovian simulation benchmarks.arXiv preprint arXiv:2603.09513, 2026

    Honghui Wang, Zhi Jing, Jicong Ao, Shiji Song, Xuelong Li, Gao Huang, and Chenjia Bai. Beyond short-horizon: Vq-memory for robust long-horizon manipulation in non-markovian simulation benchmarks.arXiv preprint arXiv:2603.09513, 2026

    arXiv 2026

  66. [66]

    Lola: Long horizon latent action learning for general robot manipulation.arXiv preprint arXiv:2512.20166, 2025

    Xiaofan Wang, Xingyu Gao, Jianlong Fu, Zuolei Li, Dean Fortier, Galen Mullins, Andrey Kolobov, and Baining Guo. Lola: Long horizon latent action learning for general robot manipulation.arXiv preprint arXiv:2512.20166, 2025

    arXiv 2025

  67. [67]

    Tacmamba: A tactile history compression adapter bridging fast reflexes and slow vla reasoning.arXiv preprint arXiv:2603.01700, 2026

    Zhenan Wang, Yanzhe Wang, Meixuan Ren, Peng Li, Yang Liu, Yifei Nie, Limin Long, Yun Ye, Xiaofeng Wang, Zhen Zhu, et al. Tacmamba: A tactile history compression adapter bridging fast reflexes and slow vla reasoning.arXiv preprint arXiv:2603.01700, 2026

    arXiv 2026

  68. [68]

    Ava-vla: Improving vision-language-action models with active visual attention

    Lei Xiao, Jifeng Li, Juntao Gao, Feiyang Ye, Yan Jin, Jingjing Qian, Jing Zhang, Yong Wu, and Xiaoyuan Yu. Ava-vla: Improving vision-language-action models with active visual attention. arXiv preprint arXiv:2511.18960, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    Affordance field intervention: Enabling vlas to escape memory traps in robotic manipulation.arXiv preprint arXiv:2512.07472, 2025

    Siyu Xu, Zijian Wang, Yunke Wang, Chenghao Xia, Tao Huang, and Chang Xu. Affordance field intervention: Enabling vlas to escape memory traps in robotic manipulation.arXiv preprint arXiv:2512.07472, 2025

    arXiv 2025

  70. [70]

    Kv-efficient vla: A method to speed up vision language models with rnn-gated chunked kv cache.arXiv preprint arXiv:2509.21354, 2025

    Wanshun Xu, Long Zhuang, and Lianlei Shan. Kv-efficient vla: A method to speed up vision language models with rnn-gated chunked kv cache.arXiv preprint arXiv:2509.21354, 2025

    arXiv 2025

  71. [71]

    Egovla: Learning vision-language-action models from egocentric human videos.arXiv preprint arXiv:2507.12440, 2025

    Ruihan Yang, Qinxi Yu, Yecheng Wu, Rui Yan, Borui Li, An-Chieh Cheng, Xueyan Zou, Yunhao Fang, Xuxin Cheng, Ri-Zhao Qiu, et al. Egovla: Learning vision-language-action models from egocentric human videos.arXiv preprint arXiv:2507.12440, 2025

    Pith/arXiv arXiv 2025

  72. [72]

    Lohovla: A unified vision- language-action model for long-horizon embodied tasks.arXiv preprint arXiv:2506.00411, 2025

    Yi Yang, Jiaxuan Sun, Siqi Kou, Yihan Wang, and Zhijie Deng. Lohovla: A unified vision- language-action model for long-horizon embodied tasks.arXiv preprint arXiv:2506.00411, 2025

    arXiv 2025

  73. [73]

    Lilo-vla: Compositional long-horizon manipulation via linked object-centric policies.arXiv preprint arXiv:2602.21531, 2026

    Yue Yang, Shuo Cheng, Yu Fang, Homanga Bharadhwaj, Mingyu Ding, Gedas Bertasius, and Daniel Szafir. Lilo-vla: Compositional long-horizon manipulation via linked object-centric policies.arXiv preprint arXiv:2602.21531, 2026. 14

    arXiv 2026

  74. [74]

    Helm: Harness-enhanced long-horizon memory for vision-language-action manipulation.arXiv preprint arXiv:2604.18791, 2026

    Zijian Zeng, Fei Ding, Huiming Yang, and Xianwei Li. Helm: Harness-enhanced long-horizon memory for vision-language-action manipulation.arXiv preprint arXiv:2604.18791, 2026

    Pith/arXiv arXiv 2026

  75. [75]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. InProceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

    2023

  76. [76]

    Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

    Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. InProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023. doi: 10.15607/RSS.2023.XIX.016

    work page doi:10.15607/rss.2023.xix.016 2023

  77. [77]

    Universal actions for enhanced embodied foundation models

    Jinliang Zheng, Jianxiong Li, Dongxiu Liu, Yinan Zheng, Zhihao Wang, Zhonghong Ou, Yu Liu, Jingjing Liu, Ya-Qin Zhang, and Xianyuan Zhan. Universal actions for enhanced embodied foundation models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22508–22519, 2025

    2025

  78. [78]

    X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model.arXiv preprint arXiv:2510.10274, 2025

    Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, et al. X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model.arXiv preprint arXiv:2510.10274, 2025

    Pith/arXiv arXiv 2025

  79. [79]

    Dexgraspvla: A vision-language-action framework towards general dexterous grasping

    Yifan Zhong, Xuchuan Huang, Ruochong Li, Ceyao Zhang, Zhang Chen, Tianrui Guan, Fanlian Zeng, Ka Nam Lui, Yuyao Ye, Yitao Liang, et al. Dexgraspvla: A vision-language-action framework towards general dexterous grasping. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 18836–18844, 2026

    2026

  80. [80]

    memory traps

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning, pages 2165–2183. PMLR, 2023. 15 A Limitations and Scope Our recurrent formulation is trained with TBPTT, so gradien...

    2023