Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Force-history memory lets a robot policy count contact events that images cannot reveal, yielding 83% average success on three manipulation tasks.

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

2026-08-01 15:37 UTC pith:3RXBZD7K

load-bearing objection A genuine new idea — force history as VLA memory — with clean ablations, but the counting mechanism is under-verified and 18 trials per cell will not carry the load. the 3 major comments →

arxiv 2607.18231 v1 pith:3RXBZD7K submitted 2026-07-20 cs.RO

FM-VLA: Force-based Memory for Vision-Language-Action Models in Contact-Rich Manipulation

Ruicheng Li , Qixiu Li , Ruichun Ma , Yu Deng , Lin Luo , Zhiying Du , Jianfeng Xiang , Huizhi Liang
show 3 more authors
Ruicheng Wang Jiaolong Yang Baining Guo
This is my paper
classification cs.RO
keywords vision-language-action modelsforce/torque sensingtemporal memoryvariational autoencodercontact-rich manipulationnon-Markovian policyflow-matching action expertrobotic manipulation
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.

The paper claims that a robot policy can solve tasks that require remembering past interactions—how many times a button was pressed, how many wiping passes remain, which cup already was lifted—by compressing the full wrist force/torque history into a compact latent representation and feeding it to a vision-language-action model. Vision alone fails because the button barely moves and the cups look identical after being replaced. The authors show that a variational autoencoder, trained only to reconstruct wrench time-series, yields tokens that the action expert can use to count and track contact events. On three contact-rich bimanual tasks the method reaches 83.3% average success, far above a memoryless baseline (27.8%) and a visual-memory baseline (53.7%), while adding only about three milliseconds of inference overhead.

Core claim

The paper's central claim is that the accumulated wrench stream is a usable temporal memory for non-Markovian manipulation. A VAE pretrained on wrench reconstruction compresses the full episode's force/torque signal into eight latent tokens; these are projected and appended to the action-expert suffix, alongside a single token summarizing the last second of joint-state history. This lets the policy count audible button clicks and full wiping rounds, and remember which visually identical cup has already been inspected. Removing either the force memory or the short state window degrades average success to roughly 26% and 41%, while both together achieve 83.3%.

What carries the argument

The load-bearing component is a frozen variational autoencoder (VAE) trained solely to reconstruct smoothed, quantile-normalized 6-axis wrench time-series. Its encoder is a cross-attention stack that maps arbitrarily long force histories to eight learned latent tokens; only the posterior mean is used at inference. A second, lightweight module projects the last ~0.9 seconds of joint positions and gripper states into a single token. Both token groups are appended after the noisy-action tokens of a flow-matching action expert, leaving the base VLA's positional layout untouched and adding negligible compute.

Load-bearing premise

The method assumes every countable contact event (button click, wipe round, cup lift) leaves a distinct, repeatable peak in the 30 Hz filtered wrist wrench stream, so that eight VAE tokens can recover the exact event count; if contacts are soft, masked by the robot's own motion, or if the audible-click criterion diverges from wrench events, counting fails.

What would settle it

Run the button task with a button whose click is silent but mechanically identical, or with force amplitudes jittered randomly across presses; if the model's count accuracy drops, then the distinct-wrench-impulse premise—not the VLA architecture—is the load-bearing element.

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

If this is right

  • VLAs can solve contact-counting and progress-tracking tasks without relying on vision memory, which is ambiguous when scenes barely change.
  • The force-memory module adds only ~3 ms of latency, making temporal conditioning feasible at real-time control rates.
  • The VAE is task-agnostic: a single pretrained encoder serves all three tasks, suggesting one wrench latent space can be reused across manipulations.
  • Neither the long-horizon force memory nor the short-horizon state window alone suffices; their combination is the actual contribution.
  • The method's success on cup-lifting shows force/torque memory can encode object-level interactions, not just contact counting.

Where Pith is reading between the lines

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

  • If the VAE latent truly encodes event counts, the same tokens could be reused as a progress signal for reinforcement learning or as a reward-shaping term, which the paper does not explore.
  • The approach should generalize to any high-rate, low-dimensional proprioceptive or tactile stream, such as joint-torque estimates or tactile-array readings, where the same reconstruction pretraining recipe may apply.
  • A silent but mechanically identical button press would fail the paper's audible-click success criterion even if the wrench signature is the same, suggesting the evaluation may conflate acoustic feedback with force events; testing with a silent button would separate the two.
  • Because the VAE is trained on the demonstration dataset, large-scale pretraining on diverse force recordings could reduce the need for task-specific data, but this remains untested.

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 FM-VLA, a vision-language-action model augmented with a force-based memory module. Long-horizon wrist force/torque histories are compressed by a frozen, reconstruction-pretrained Perceiver-IO VAE into eight latent tokens, which are appended to the action-expert suffix together with a short window of joint-state history. The VLA is fine-tuned with a flow-matching objective on three contact-rich, memory-dependent bimanual tasks: locating a block under two cups, pressing a button a specified number of times, and wiping a bowl a specified number of rounds. The paper reports an average success rate of 83.3%, substantially exceeding memoryless, short-window force, and vision-memory baselines, with only 3.3 ms added inference latency. Ablations indicate that both the force memory and the short state history contribute, and that the VAE pretraining outperforms GRU and Q-Former alternatives.

Significance. If the central claim holds, FM-VLA is a meaningful advance: it addresses a real limitation of current VLAs—the inability to reason over non-Markovian interaction history—with a compact, low-overhead modality that is naturally suited to contact-rich tasks and is complementary to visual memory. The paper is careful in several respects: it provides detailed hyperparameters, describes baseline reimplementations precisely, reports per-task trial counts, and measures inference latency. The task suite is well chosen: the button and wiping tasks are genuinely ambiguous in vision and depend on contact-event counting. The main weakness is that the paper never directly demonstrates that the VAE latent actually encodes the contact-event count that the success criterion requires; the evidence is task-level success alone, which can be explained by alternative cues such as episode duration, language-specified count, or the policy's own action history. Given that the proposed mechanism is the paper's core contribution, this missing evidence is load-bearing and should be addressed before the claims can be accepted.

major comments (3)
  1. [§3.2.2, §3.3, Eq. (2), Eq. (4), Table 1] The central claim is that the frozen VAE's 8-token summary encodes completed contact-event counts. The paper asserts this in §3.2.2 ('contact counts') and in §4.3 ('the latent space to encode ... contact counts'), but no direct evidence is provided. A reconstruction objective with free-bits KL does not guarantee count separability. The authors should probe the posterior mean μ_f, e.g., with a linear classifier on histories containing 0/1/2/3 events, or by visualizing latent trajectories during a single episode. Without such a probe, the button/wipe gains could be carried by episode length, elapsed time, or the policy's own action history rather than by the force-memory content.
  2. [§3.2.1, §3.2.3] The paper states that random noise pre-padding is used during training to remove the sequence-length shortcut, but that 'pre-padding is disabled at inference.' At inference, the input length is exactly the episode length, and the absolute Fourier positional encoding preserves absolute position. Combined with the language instruction specifying N and the short state window, the policy could in principle count by duration or by its own executed actions. To rule out this shortcut, the authors should include controls that decorrelate duration from count (variable inter-event timing, random delays, faster/slower demonstrations) and ideally ablate the force tokens entirely while keeping the state window and language input. The current experiments do not exclude this alternative explanation.
  3. [§4.1, Table 1, §4.3] All success rates are based on 18 trials per task, and some key comparisons are not statistically robust. For example, on Wipe Dishes the VAE model achieves 14/18 (77.8%) versus the Q-Former's 10/18 (55.6%); this difference is not significant at the 5% level (exact binomial/Fisher test p≈0.29), despite being used to support the architecture-ablation claim. The paper should report confidence intervals or exact tests, and ideally evaluate over multiple seeds or checkpoints. The token-count ablation (Fig. 3) is also based on a single task with no error bars.
minor comments (4)
  1. [§4.3] The claim that the Q-Former 'overfits to instantaneous peaks instead of the holistic temporal structure' is presented without supporting analysis. If this is an interpretation of the ablation result, it should be labeled as such, or supported by inspecting the learned Q-Former attention.
  2. [§3.2.2 / Eq. (4)] The free-bits term is described as a per-dimension KL floor, but the equation reads max(D_KL, λ) and is averaged over K·d_z dimensions. Please clarify whether the max is applied per dimension before averaging, and state explicitly how the gradient is stopped on dimensions below the floor.
  3. [Table 2 and Figure 3] Latency values are reported with ± but no number of repeated measurements or hardware details beyond the GPU model. The token-count ablation reports success rates over a single task; a brief note on variability would help.
  4. [Throughout] Minor typographical issues: 'π0.5' is sometimes written as 'π 0.5'; Eq. (5) contains stray carriage-return artifacts; '3 ms' appears as '3 ms' in Table 2 caption. These should be cleaned up.

Circularity Check

0 steps flagged

No circular derivation: the force-memory VAE and downstream policy are trained with separate objectives, and the central claims are empirical rather than definitional.

full rationale

FM-VLA is an empirical robotics paper, not a derivation from first principles. The claimed chain is: pretrain a VAE on wrench reconstruction (Eq. 4), freeze it, inject its posterior-mean tokens plus a short state token into a flow-matching action expert (Eq. 5), and measure task success. None of these steps defines the target result in terms of its inputs: the VAE is trained only on reconstruction with free-bits KL and no count/task labels, and the policy is trained with the standard rectified-flow action objective. The assertion that the latent encodes contact counts (§3.2.2, §4.3) is an empirical hypothesis, not a tautology; it could fail or be supported by other cues. The paper explicitly identifies the episode-length shortcut concern and randomizes pre-padding during training (§3.2.1), although disabling it at inference leaves a possible alternative cue; that is an experimental-control gap, not circularity. The K=8 token count is selected by peaking on Wipe (Fig. 3), so the reported Wipe success is partly model-selected; this is a statistical overfitting caveat, but the main VAE-vs-Q-Former comparison holds K fixed and therefore provides independent evidence. Self-citations such as CogACT [16] appear only in related-work lists and are not load-bearing; no uniqueness theorem or ansatz is imported from author prior work. The limitations section admits only capacity/scale limitations, not a circular dependency. Overall, no step in the paper's argument reduces by construction to its inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The ledger is dominated by design choices (K, β, λ, α, σ, window length) and domain assumptions about wrench-event separability, not by free physics-like parameters. The single most consequential choice is K=8, selected by ablation on the Wipe task that is itself included in the headline average. The methodological assumptions that most affect verdict strength are the faithful-reimplementation assumption for π-MEM/TA-VLA and the assumption that the audible-click criterion coincides with wrench-countable events. No new physical entities are postulated; 'force memory tokens' are learned model components, so the graviton-problem test does not apply.

free parameters (5)
  • Force memory token count K = 8
    Chosen as the peak of the Wipe-task token ablation (Figure 3: 11.1/77.8/55.6/33.3 for K=4/8/16/32) and then used for the headline 83.3% average; tuned on the same task that is part of the reported average.
  • VAE KL weight β and free-bits floor λ = β=1e-3, λ=0.5 nats
    Hand-chosen in Table 3; controls how compressible the force latents are; not derived from any principle.
  • EMA smoothing α and noise-pad σ = α=0.3, σ=0.05
    Hand-chosen preprocessing (App. E.1); α sets the tradeoff between noise removal and event-onset preservation — the exact quantity the counting task depends on.
  • Short state window (taps, stride) = 10 taps at stride 3 (~0.9 s)
    Hand-picked (App. E.1); window length controls how much pre-contact motion context the policy gets.
  • Wrench quantile normalization limits = dataset q01/q99
    Per-dataset normalization statistics (Table 4) are fit to the full collected wrench data; the counting behavior inherits these fitted scales.
axioms (6)
  • standard math Rectified-flow (flow-matching) training objective and VAE ELBO with reparameterization are valid for this setting; RoPE positions of suffix tokens are preserved.
    Eqs. (4)-(5) in §3.3 and §3.2.3 rely on standard flow-matching/ELBO machinery inherited from π0.5 and the VAE literature [9,10,37,38]; not re-derived here.
  • domain assumption 30 Hz right-wrist wrench (downsampled from 100 Hz, EMA-smoothed with α=0.3) preserves unambiguous signatures of each contact event.
    §3.2.1 and Task 2 in §4.1; the whole counting mechanism depends on wrench impulses being distinguishable and countable.
  • domain assumption Success criteria (audible-click count, full back-and-forth wipe rounds, no cup re-inspection) are measurable, reliable ground truth guiding training and evaluation.
    App. B.2 defines them; if the criteria are noisy or gameable (e.g., gripper-opening timing), success rates overstate the counting ability.
  • domain assumption The π-MEM and TA-VLA reimplementations (App. D) faithfully capture the published methods' behavior.
    App. D provides detailed reimplementations; no validation against official code or reported numbers is given, so the baseline comparisons depend on reimplementation fidelity.
  • ad hoc to paper Random Gaussian noise prepadding (σ=0.05, up to 10 s) prevents the policy from using sequence length as a temporal-progress shortcut and does not harm inference.
    §3.2.1; a paper-specific augmentation whose efficacy is asserted rather than ablated.
  • domain assumption The VAE posterior mean (frozen, eval mode, no sampling) retains the event-count information needed by the policy.
    §3.3 'Only the posterior mean μ_f is used'; assumes the VAE latent compresses force history without collapsing task-relevant count structure.

pith-pipeline@v1.3.0-alltime-deepseek · 14102 in / 22379 out tokens · 172876 ms · 2026-08-01T15:37:27.207004+00:00 · methodology

0 comments
read the original abstract

Vision-language-action (VLA) models have achieved impressive generalization in robotic manipulation, and recent memory-augmented VLAs have relaxed the Markovian assumption by conditioning on past images or language summaries. Vision-based memory approaches address this by conditioning on sampled past image frames, but they are computationally expensive and fundamentally limited when temporal events are visually ambiguous, e.g., pushing a button multiple times with small movements. We propose FM-VLA, a VLA model with force-based memory, enabling temporal context reasoning for non-Markovian, contact-rich manipulation. We encode force histories into compact force memory tokens with a variational autoencoder (VAE) pretrained with force time series reconstruction. By projecting force latent representations and short state history as additional conditioning tokens to the action expert module, we enable VLAs to leverage accumulated contact event history to guide manipulation. We evaluate FM-VLA on three memory-dependent tasks, including finding a hidden block, pressing a button, and wiping a dish for a specific number of times. Our lightweight force memory achieves over 80% success rate with minimal inference overhead, significantly outperforming baseline approaches. Project page: https://qft-333.github.io/FM-VLA-Page/

Figures

Figures reproduced from arXiv: 2607.18231 by Baining Guo, Huizhi Liang, Jianfeng Xiang, Jiaolong Yang, Lin Luo, Qixiu Li, Ruicheng Li, Ruicheng Wang, Ruichun Ma, Yu Deng, Zhiying Du.

Figure 1
Figure 1. Figure 1: Comparison between visual memory based VLA and FM-VLA, which incorporates force (wrench) based memory to enable temporal context understanding for non-Markovian, contact-rich manipulation tasks. Abstract: Vision-language-action (VLA) models have achieved impressive gen￾eralization in robotic manipulation, and recent memory-augmented VLAs have relaxed the Markovian assumption by conditioning on past images … view at source ↗
Figure 2
Figure 2. Figure 2: Overview. We augment a VLA with lightweight force-based temporal memory. Force/torque histories are encoded by a VAE encoder into latent representations, projected via MLP into force memory tokens that condition the flow matching action expert. A short-term state memory token is further appended to provide motion context. • We introduce a lightweight VAE to compress long-horizon force signals into a compac… view at source ↗
Figure 3
Figure 3. Figure 3: Token count ablation. 1 1 2 3 4 5 6 7 1 2 3 4 5 6 7 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Trajectory and force signal visualization. For each task, we visualize the FM-VLA inference trajectory and a selected channel of force readings, with each frame’s time marked. FM￾VLA correctly memorizes contact events and complete manipulation successfully. 4.4 Inference Efficiency We compare FM-VLA with the base policy π0.5 and vision-based memory in terms of inference la￾tency on an RTX 4090, as shown in… 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

42 extracted references · 28 linked inside Pith

  1. [1]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023.https://arxiv.org/abs/2307.15818

    Pith/arXiv arXiv 2023

  2. [2]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, et al. OpenVLA: An open-source vision-language-action model. InConference on Robot Learning (CoRL), 2024. arXiv:2406.09246

    Pith/arXiv arXiv 2024

  3. [3]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, et al.π 0: A vision-language-action flow model for general robot control. InRobotics: Science and Systems (RSS), 2025. arXiv:2410.24164

    Pith/arXiv arXiv 2025

  4. [4]

    Black, N

    Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, et al.π 0.5: A vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H. Fan, X. Zhang, et al. MemoryVLA: Perceptual-cognitive memory in vision-language-action models for robotic manipulation. In International Conference on Learning Representations (ICLR), 2026. arXiv:2508.19236

    Pith/arXiv arXiv 2026

  6. [6]

    Torne, K

    M. Torne, K. Pertsch, H. Walke, K. Vedder, S. Nair, B. Ichter, A. Z. Ren, H. Wang, J. Tang, K. Stachowicz, K. Dhabalia, M. Equi, Q. Vuong, J. T. Springenberg, S. Levine, C. Finn, and D. Driess. MEM: Multi-scale embodied memory for vision language action models.arXiv preprint arXiv:2603.03596, 2026

    arXiv 2026

  7. [7]

    J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y . Song, P. Cai, C. Lu, and W. Zhang. ForceVLA: Enhancing vla models with a force-aware MoE for contact- rich manipulation. InAdvances in Neural Information Processing Systems (NeurIPS), 2025. arXiv:2505.22159

    arXiv 2025

  8. [8]

    Zhang, H

    Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao. TA-VLA: Elucidating the design space of torque-aware vision-language-action models. InConference on Robot Learning (CoRL), 2025. arXiv:2509.07962

    Pith/arXiv arXiv 2025

  9. [9]

    P. K. Diederik and W. Max. An introduction to variational autoencoders.Foundations and Trends® in Machine Learning, 12(4):307–392, 2019

    2019

  10. [10]

    Jaegle, S

    A. Jaegle, S. Borgeaud, J.-B. Alayrac, C. Doersch, C. Ionescu, D. Ding, S. Koppula, D. Zoran, A. Brock, E. Shelhamer, O. H ´enaff, M. M. Botvinick, A. Zisserman, O. Vinyals, and J. Car- reira. Perceiver IO: A General Architecture for Structured Inputs & Outputs, 2022. URL https://arxiv.org/abs/2107.14795

    Pith/arXiv arXiv 2022

  11. [11]

    Steiner, A

    A. Steiner, A. S. Pinto, M. Tschannen, D. Keysers, X. Wang, Y . Bitton, A. Gritsenko, M. Min- derer, A. Sherbondy, S. Long, et al. Paligemma 2: A family of versatile vlms for transfer.arXiv preprint arXiv:2412.03555, 2024. 9

    Pith/arXiv arXiv 2024

  12. [12]

    Comanici, E

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen, et al. Gemini 2.5: Pushing the frontier with advanced rea- soning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu. RDT-1B: A diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Gemini Robotics: Bringing ai into the physical world.arXiv preprint arXiv:2503.20020, 2025

    Gemini Robotics Team. Gemini Robotics: Bringing ai into the physical world.arXiv preprint arXiv:2503.20020, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Q. Li, Y . Liang, Z. Wang, L. Luo, X. Chen, M. Liao, F. Wei, Y . Deng, S. Xu, 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

  17. [17]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    D. Qu, H. Song, Q. Chen, Y . Yao, X. Ye, Y . Ding, Z. Wang, J. Gu, B. Zhao, D. Wang, et al. Spatialvla: Exploring spatial representations for visual-language-action model.arXiv preprint arXiv:2501.15830, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Cheang, S

    C. Cheang, S. Chen, Z. Cui, Y . Hu, L. Huang, T. Kong, H. Li, Y . Li, Y . Liu, X. Ma, et al. Gr-3 technical report.arXiv preprint arXiv:2507.15493, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    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

    arXiv 2026

  21. [21]

    M. Lin, X. Liang, B. Lin, L. Jingzhi, Z. Jiao, K. Li, Y . Sun, W. Liufu, Y . Ma, Y . Liu, S. Zhao, Y . Zhuang, and X. Liang. EchoVLA: Synergistic declarative memory for vla-driven mobile manipulation.arXiv preprint arXiv:2511.18112, 2025

    arXiv 2025

  22. [22]

    Y . Lei, Z. Liang, H. Zhang, and P. Luo. VPWEM: Non-markovian visuomotor policy with working and episodic memory.arXiv preprint arXiv:2603.04910, 2026

    Pith/arXiv 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

    arXiv 2026

  24. [24]

    J. Sun, B. Yang, J. Zhang, N. Ma, C. Wu, S. Zhang, Y . Huang, Q. Wang, S. Liang, and Y . Chen. TempoFit: Plug-and-play layer-wise temporal kv memory for long-horizon vision-language- action manipulation.arXiv preprint arXiv:2603.07647, 2026

    arXiv 2026

  25. [25]

    Haresh, D

    S. Haresh, D. Dijkman, A. Bhattacharyya, and R. Memisevic. Notes-to-self: Scratchpad aug- mented VLAs for memory dependent manipulation tasks. InIEEE International Conference on Robotics and Automation (ICRA), 2026. arXiv:2602.21013

    Pith/arXiv arXiv 2026

  26. [26]

    Z. Li, B. Hu, R. Shao, G. Chen, D. Jiang, P. Xie, J. Hao, and L. Nie. Global prior meets local consistency: Dual-memory augmented vision-language-action model for efficient robotic manipulation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

  27. [27]

    Chung, T

    N. Chung, T. Hanyu, T. Nguyen, H. Le, F. Bumgarner, D. M. H. Nguyen, K. V o, K. Yamazaki, C. Rainwater, T. Kieu, A. Nguyen, and N. Le. Rethinking progression of memory state in robotic manipulation: An object-centric perspective. InAAAI Conference on Artificial Intelli- gence, 2026. arXiv:2511.11478. 10

    arXiv 2026

  28. [28]

    Zhang, Y

    S. Zhang, Y . Yuan, H. Tang, Y . Zheng, Q. Yu, P. Li, G. Huang, H. Huang, X. Quan, and J. Hao. ForceFlow: Learning to feel and act via contact-driven flow matching.arXiv preprint arXiv:2605.11048, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    Zhang, H

    K. Zhang, H. Zhang, Z. Xu, Z. Zhang, M. R. I. Prince, X. Li, X. Han, Y . Zhou, A. Ajoudani, and Y . She. TacVLA: Contact-aware tactile fusion for robust vision-language-action manipulation. arXiv preprint arXiv:2603.12665, 2026

    arXiv 2026

  30. [30]

    Gubernatorov, M

    K. Gubernatorov, M. Sannikov, I. Mikhalchuk, E. Kuznetsov, M. Artemov, O. F. Ouwatobi, M. Fernando, A. Asanov, Z. Guo, and D. Tsetserukou. HapticVLA: Contact-rich manipula- tion via vision-language-action model without inference-time tactile sensing.arXiv preprint arXiv:2603.15257, 2026

    arXiv 2026

  31. [31]

    X. Li, M. Cai, J. Xu, J. Zhu, H. Fan, Y . Shen, G. Ren, and H. Dong. AT-VLA: Adaptive tactile injection for enhanced feedback reaction in vision-language-action models.arXiv preprint arXiv:2605.07308, 2026

    Pith/arXiv arXiv 2026

  32. [32]

    Huang, P

    Y . Huang, P. Lin, W. Li, D. Li, J. Li, J. Jiang, C. Xiao, and Z. Jiao. TaF-VLA: Tactile- force alignment in vision-language-action models for force-aware manipulation.arXiv preprint arXiv:2601.20321, 2026

    arXiv 2026

  33. [33]

    Zhang, P

    C. Zhang, P. Hao, X. Cao, X. Hao, S. Cui, and S. Wang. VTLA: Vision-tactile-language-action model with preference learning for insertion manipulation.arXiv preprint arXiv:2505.09577, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    G. Ye, Z. Zhang, X. Zhao, S. Wu, H. Lu, S. Lu, and H. Liu. Learning to feel the future: DreamTacVLA for contact-rich manipulation.arXiv preprint arXiv:2512.23864, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Beyer, A

    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

    Pith/arXiv arXiv 2024

  36. [36]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre- training.arXiv preprint arXiv:2303.15343, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Lipman, R

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Pith/arXiv arXiv 2022

  39. [39]

    Agibot g1 humanoid robot.https://www.agibot.com/products/G1,

    Zhiyuan Robotics. Agibot g1 humanoid robot.https://www.agibot.com/products/G1,

  40. [40]

    Chung, C ¸

    J. Chung, C ¸ . G¨ulc ¸ehre, K. Cho, and Y . Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling.CoRR, abs/1412.3555, 2014. URLhttp://arxiv.org/ abs/1412.3555

    Pith/arXiv arXiv 2014

  41. [41]

    J. Li, D. Li, S. Savarese, and S. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.arXiv preprint arXiv:2301.12597, 2023. 11 A Additional Real-World Results Additional real-world inference results of our method can be found in the supplementary video. B Task Definitions and Evaluation Protocol Thi...

    Pith/arXiv arXiv 2023

  42. [2024]

    Accessed: 2026-05-24

    2026