Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

MemoAct: a two-tier robot memory that tracks every step and still recalls the start, outperforming the strongest baseline by 24.5% on its own benchmark.

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

2026-08-04 05:43 UTC pith:WTSHZMRY

load-bearing objection A plausible dual-memory architecture with honest self-assessment; the headline gains need error bars and official baselines before the claims are fully convincing. the 4 major comments →

arxiv 2603.18494 v2 pith:WTSHZMRY submitted 2026-03-19 cs.RO

MemoAct: Atkinson-Shiffrin-Inspired Hierarchical Memory-Augmented Policy for Robotic Manipulation

Liufan Tan , Jiale Li , Gangshan Jing This is my paper
classification cs.RO
keywords memory-augmented policyvisuomotor policyAtkinson-Shiffrin modelhierarchical memorytask state trackinglong-horizon retentionrobotic manipulationdiffusion policy
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 proposes MemoAct, a robot policy with a three-tier memory modeled on the Atkinson–Shiffrin theory of human memory. Its central idea is that a robot facing memory-dependent manipulation needs two kinds of historical storage at once: a lossless short-term bank that keeps recent frames intact for precise tracking of subtask state, and a compressed long-term bank that keeps gist-level records of older events for robust recall. MemoAct reports large gains over the strongest baseline — 24.5% on the new MemoryRTBench, 21.0% on RMBench, and 22.5% in real-world trials. A reader should care because existing history-aware policies force a tradeoff between tracking the current step and remembering the initial state, and MemoAct argues that the tradeoff is avoidable with hierarchical consolidation.

Core claim

MemoAct's claim is that a single policy can simultaneously track fine-grained task state and retain long-horizon context by separating memory into a lossless short-term bank (capacity 6) and a compressed long-term bank (capacity 8). Each new observation is distilled into one token via a learnable readout over DINOv2 patch features; this sensory memory is appended to the short-term bank, and when the bank fills, the oldest three tokens are compressed through a causal transformer encoder into a summary token that migrates to the long-term bank. When the long-term bank fills, the two most similar adjacent tokens are merged instead of evicted FIFO-style. A gated cross-attention step fuses retrie

What carries the argument

The load-bearing component is the long short-term memory consolidation module, a joint update rule for the two banks. The short-term bank (Ts=6) stores lossless sensory tokens so the policy can tell apart perceptually identical moments; the long-term bank (Tk=8) stores compressed tokens created by a causal transformer's summary token, preserving older context beyond the short-term window. When long-term capacity runs out, adjacent tokens with highest pairwise similarity are merged, extending the horizon without a hard FIFO cutoff. These two mechanisms together — not the action decoder or the visual backbone — are what the paper credits for simultaneous tracking and retention.

Load-bearing premise

The assumption that carries the most weight is that squeezing an entire RGB frame into a single token still leaves enough spatial detail for precise task-state tracking; the paper concedes this compression 'inevitably compromises visual fidelity' and the method scores only 4% on the spatial recall task Observe and Pick Up.

What would settle it

Scale up spatial precision demand in a controlled way: in a 'Put Back Block' style task, rotate the block by 90 degrees after the first observation and require the robot to restore the original orientation after a delay. If the single-token sensory memory cannot distinguish orientations, success will collapse toward the 4% seen on Observe and Pick Up; if MemoAct retains near-perfect success, the spatial-compression concern is not the limiting factor. A cheaper check is an ablation that replaces the one-token readout with a 4-token readout and measures whether fine-grained recall tasks improve.

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

If this is right

  • Memory-dependent manipulation can be handled without choosing between FIFO windows and lossy compression; keeping recent frames lossless and compressing only older tokens retains both precision and horizon.
  • The memory consolidation module transfers as a plug-in: adding it to the point-cloud policy DP3 raises average success on MemoryRTBench from 26% to 76.5%.
  • Bank capacities are not free parameters to maximize: setting long-term capacity to 12 lowers average success to 91%, indicating that excessive history injects noise rather than help.
  • Tasks that only need short-horizon state tracking (like sequential hammer tap) depend mainly on the short-term bank, while tasks requiring initial-state recall (like put-back-block) depend on the long-term bank.

Where Pith is reading between the lines

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

  • The one-token sensory bottleneck suggests an immediate variant: replacing the single learnable readout with a small set of spatially-indexed tokens should improve fine-grained recall tasks (e.g., RMBench's Observe and Pick Up) without changing the memory banks.
  • Because the long-term bank merges the most similar adjacent tokens, it implicitly implements a recency-biased forgetting curve; an explicit test would compare it against event-salience merging, where tokens are merged based on task relevance rather than similarity.
  • The reported dependence on temporal positional embeddings hints that order information is what allows the policy to disambiguate identical observations; this predicts that any task with re-identifiable cues but shuffled order would break the method.

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

4 major / 5 minor

Summary. The paper proposes MemoAct, a memory-augmented visuomotor policy inspired by the Atkinson–Shiffrin model. It consists of a sensory distillation module that compresses each RGB image and proprioceptive state into a single token (Eq. 1), a hierarchical memory module with a lossless short-term bank (capacity Ts=6), a compressed long-term bank (capacity Tk=8) updated by causal-attention consolidation and similarity-based merging, and a conditional diffusion action decoder. The authors introduce MemoryRTBench (four simulation tasks from RoboTwin 2.0) to probe sequential, spatial, and episodic memory, and also evaluate on RMBench and two real-world tasks. They report large success-rate improvements over Markovian baselines (ACT, DP) and two memory-mechanism baselines (SAMP, MVMP), specifically a 24.5% average gain over MVMP on MemoryRTBench, 21.0% on RMBench, and 22.5% in real-world tasks, and claim this resolves the tradeoff between FIFO windowing and lossy compression.

Significance. If the reported results withstand rigorous evaluation, MemoAct would be a useful contribution: the two-tier memory design with lossless short-term and compressed long-term banks is a clean instantiation of a known cognitive principle, the plug-in extension to DP3 (Table V) suggests practical generality, and the proposed MemoryRTBench could help standardize evaluation of history-aware policies. The ablation study (Table IV) and capacity analysis (Fig. 5) provide initial evidence that the components matter. However, the significance is currently conditional on evaluation choices that are not fully controlled: the headline margins are obtained after selecting memory capacities on the same benchmark used for the final comparison, the strongest baselines are author re-implementations rather than official models, and no variance information is reported anywhere. The paper also explicitly acknowledges a spatial-fidelity limitation that narrows the claimed scope.

major comments (4)
  1. [Section IV.C, Fig. 5] The memory capacities Ts=6, Tk=8, and Nsc=3 are selected using the same MemoryRTBench tasks on which the headline result in Table I is reported. The text states: 'Thus, we set Tk=8, Ts=6, and Nsc=3 to optimally balance performance and efficiency.' This is a model-selection loop: the reported 96.5% average is the best among the evaluated capacity configurations, while the SAMP and MVMP baselines are reported as single configurations with no equivalent tuning described. The claim that MemoAct 'significantly outperforms the strongest baseline, MVMP, by 24.5%' is therefore not an unbiased comparison. The authors should fix capacities before evaluation on an untouched set, or report the full capacity sweep for all baselines; at minimum, they should state which of the reported numbers correspond to the selected configuration.
  2. [Section IV.A, Tables I-III] The evaluation protocol uses 50 trials per MemoryRTBench task, 100 trials per RMBench task, and 10 trials for real-world tasks, but no error bars, confidence intervals, or multiple training seeds are reported. Table II contains exact ties (Swap T 55%-55%; Observe and Pick Up 4%-4%) and large gaps (SHT 86% vs 44%), yet without interval estimates or per-seed data it is impossible to determine which differences are statistically meaningful. The ablation table (Table IV) and the capacity figure (Fig. 5) have the same problem. The authors should report mean±std or Wilson intervals over at least 3–5 training seeds, and specify the number of seeds per configuration.
  3. [Section IV.A, Baselines] The strongest baselines, MVMP and SAMP, are author re-implementations: the paper says the memory mechanisms of MemoryVLA and SAM2Act are 'extracted' while all other components remain identical to MemoAct. This controls for architecture, but it makes the 'significantly outperforms the strongest baseline' claim dependent on the fidelity of those re-implementations. No code, hyperparameters, or validation of the reimplementations against the official models is provided, and no comparison with official MemoryVLA or SAM2Act is given. If the reimplementations are weaker than the originals, the 24.5% margin overstates the benefit. The authors should either evaluate official models where feasible, release the reimplementation code, or clearly frame the comparison as being against their own memory-module baselines, and include a sensitivity analysis of baseline hyperparameters.
  4. [Section III.A, Eq. (1); Section V; Table II] The central claim that MemoAct simultaneously provides 'precise task-state tracking' and 'robust long-horizon retention' is too broad given the single-token compression of the entire RGB image. The paper's own conclusion states this 'inevitably compromises visual fidelity,' and Table II shows that on 'Observe and Pick Up,' a task requiring precise spatial recall, MemoAct achieves only 4%, matching MVMP, and ties at 55% on 'Swap T.' Thus the claimed resolution of the window-vs-compression tradeoff is only demonstrated for tasks where a one-token visual summary suffices. The authors should restrict the claim accordingly, or add a multi-token/adaptive-compression variant and evaluate it on fine-grained spatial-recall tasks.
minor comments (5)
  1. [Section II.A] Typo: 'have have greatly advanced' should be 'have greatly advanced.'
  2. [Abstract vs. Section IV.A] The abstract states MemoryRTBench comprises '6 manipulation tasks,' while Section IV.A says it 'consists of four simulation tasks' plus two real-world tasks. Please clarify whether the count includes real-world tasks.
  3. [Algorithm 1] The streaming training loop uses variables start/end that are incremented inside the trajectory loop and reset per epoch. It would help to state explicitly that each batch is a contiguous chunk of one trajectory, and to describe how the memory bank is cleared for each episode during training.
  4. [Table II] In 'Observe and Pick Up,' all methods are within 2–4%, and MemoAct ties MVMP at 4%. This is a useful illustration of the scope limitation, but it is not discussed as a tie in the text; consider acknowledging it explicitly.
  5. [Section IV.D] The DP3+MCM result is promising, but Table V reports only a single average per task without variance or training-seed details. Please add the same statistical reporting used for the main tables.

Circularity Check

0 steps flagged

No significant circularity: MemoAct's architecture and reported gains are empirical, with no derivation that reduces to its own inputs.

full rationale

The paper does not contain a derivation chain in which a predicted quantity is defined in terms of the same quantity or in which fitted parameters are renamed as predictions. The memory capacities Ts=6, Tk=8, and Nsc=3 are selected via ablation on MemoryRTBench (Fig. 5) and then used in the final evaluation on the same benchmark; this is a model-selection concern that can inflate reported numbers, but it is not circular in the sense of the taxonomy because the success rates are empirical measurements, not algebraic consequences of the chosen capacities. The MVMP and SAMP baselines are author-constructed re-implementations of memory mechanisms from other papers for controlled comparison; this raises questions of comparability and implementation fidelity, but it is not circularity. The Atkinson–Shiffrin inspiration is an external cognitive-science model, and no load-bearing claim is justified by a self-citation or by an imported uniqueness theorem. The paper explicitly acknowledges a limitation of its single-token sensory compression (Section V), which is an honest assessment rather than a circular step. Overall, the central claim is an empirical architecture comparison against external benchmarks and real-world tasks, so no circularity is present.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The approach relies on the self-created MemoryRTBench and re-implemented baselines. The main free parameters are memory capacities tuned on the same benchmark used for reporting. The one-token visual compression is admitted by the authors as a limitation, and the similarity-merging mechanism is borrowed from MemoryVLA despite being criticized as a source of failure in that same model.

free parameters (3)
  • Short-term memory capacity T_s = 6
    Selected from the ablation in Fig. 5; T_s=2 degrades SHT to 46%, and T_s=6 is chosen as the balance. Tuned on the same benchmark used for final evaluation.
  • Long-term memory capacity T_k = 8
    Selected from the ablation in Fig. 5; T_k=0 severely hurts BPR/DA-STR/DA-ITR and T_k=12 drops the average to 91%. Chosen on the benchmark's results.
  • Consolidation chunk size N_sc = 3
    Number of oldest STMB tokens compressed per consolidation step; the paper states it is set to 'optimally balance performance and efficiency' based on the same benchmark.
axioms (6)
  • domain assumption Perceptual aliasing causes Markovian policies to fail probabilistically (citations [9],[12]-[15]).
    This motivates the entire memory design and is presented as established in the Introduction.
  • domain assumption The Atkinson–Shiffrin three-tier memory model transfers from human cognition to robotic policy architecture.
    The design inspiration; no evidence beyond analogy is provided (Section I).
  • domain assumption RoboTwin 2.0 simulation is a valid proxy for real-world robotic performance.
    Most results come from the simulator; real-world validation uses only two tasks with 10 trials each (Table III).
  • domain assumption Similarity-based merging in the LTMB does not discard task-critical distinctions.
    The LTMB merges adjacent similar tokens (Section III.B); the paper criticizes MVMP for exactly this failure mode, and MemoAct inherits the mechanism.
  • domain assumption Streaming training with strict chronological order is sufficient for temporal consistency.
    Introduced in Section III.D without comparison to shuffled training or other curricula.
  • domain assumption Frozen DINOv2 features provide sufficient visual representation for manipulation.
    The backbone is frozen (Section III.A); the ablation only compares with ResNet-18, not other backbones.

pith-pipeline@v1.3.0-alltime-deepseek · 10169 in / 10077 out tokens · 96493 ms · 2026-08-04T05:43:52.364781+00:00 · methodology

0 comments
read the original abstract

Memory-augmented robotic policies are essential in handling memory-dependent tasks. However, existing approaches typically rely on simply extending the observation window, struggling to simultaneously achieve precise task-state tracking and robust long-horizon retention. To overcome these challenges, inspired by the Atkinson--Shiffrin memory model, we propose MemoAct, a hierarchical memory-augmented policy that leverages distinct memory tiers to tackle specific bottlenecks. Specifically, sensory memory filters immediate perceptual inputs, lossless short-term memory supports precise task-state tracking, and compressed long-term memory facilitates robust long-horizon retention. To enrich the evaluation landscape, we construct MemoryRTBench based on RoboTwin 2.0, comprising 6 manipulation tasks that systematically evaluate policy memory capabilities across three dimensions: sequential, spatial, and episodic memory. Extensive experiments across simulated and real-world scenarios demonstrate that MemoAct achieves superior performance compared to both existing Markovian baselines and history-aware policies. The project page is available at https://tlf-tlf.github.io/MemoActPage/.

Figures

Figures reproduced from arXiv: 2603.18494 by Gangshan Jing, Jiale Li, Liufan Tan.

Figure 1
Figure 1. Figure 1: (a) An example of a memory-dependent task. (b) Policies lacking historical awareness fail under identical [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MemoAct architecture. The sensory distillation module first encodes RGB images and proprioceptive states into high-fidelity features, termed sensory memory. This memory serves as a query to retrieve relevant historical context from the long short-term memory bank, which is processed by a temporal transformer encoder. Subsequently, a gating network adaptively fuses the retrieved history with the… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Long Short-Term Mem￾ory Consolidation Module. Newly generated observa￾tion embeddings are first appended to the STMB. Upon saturation of the STMB capacity, the earliest Nsc entries are compressed by feeding them, alongside a learnable summary token, into the temporal transformer encoder. The resulting summary token is migrated to the LTMB, while the original Nsc entries are discarded. T… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of simulation and real-world tasks. The figure illustrates four simulation tasks from MemoryRTBench [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reducing long-term capacity (Tk = 0) severely degrades tasks requiring initial state recall (BPR, DA￾STR, DA-ITR). Conversely, blindly expanding it (Tk = 12) drops the average success rate to 91%, as excessive history introduces noisy and redundant features that interfere with decision-making. For short-term memory, reducing capacity drastically impairs the SHT task (dropping to 46% at Ts = 2) and causes r… view at source ↗

discussion (0)

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

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ChainVLA: Chaining Vision-Language-Action Queries through a Unified Execution State for Long-Horizon Manipulation

    cs.RO 2026-08 conditional novelty 6.0

    ChainVLA chains vision-language-action queries by carrying both retrieved task progress and the previous prediction's unexecuted suffix into each new action horizon, reaching 62.8% on RMBench and 98.8% on LIBERO.

  2. Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation

    cs.RO 2026-06 unverdicted novelty 6.0

    Chronos elevates full observation history to the policy's latent state via selective SSM tokens and a Schrödinger-inspired acceleration bridge, achieving large gains on memory-dependent robot tasks with fewer parameters.

  3. EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies

    cs.CV 2026-06 unverdicted novelty 6.0

    EventVLA introduces foundational visual anchors and a Keyframe Evidence Memory module that predicts future keyframe probabilities from VLA embeddings to improve long-horizon task success by an average of 40% on 17 sim...

  4. EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies

    cs.CV 2026-06 unverdicted novelty 5.0

    EventVLA uses a foresight-driven Keyframe Evidence Memory module to selectively store sparse task-critical visual events predicted from VLA latent embeddings, reporting +40% average success rate gains on 17 simulation...

Reference graph

Works this paper leans on

38 extracted references · 15 linked inside Pith · cited by 3 Pith papers

  1. [1]

    π ∗ 0.6: a vla that learns from experience,

    P. Intelligence, A. Amin, R. Aniceto, A. Balakrishna, K. Black, K. Conley, G. Connors, J. Darpinianet al., “π ∗ 0.6: a vla that learns from experience,” 2025. [Online]. Available: https://arxiv.org/abs/2511.14759

    Pith/arXiv arXiv 2025

  2. [2]

    Rdt2: Enabling zero-shot cross-embodiment gener- alization by scaling up umi data,

    R. Team, “Rdt2: Enabling zero-shot cross-embodiment gener- alization by scaling up umi data,” 2025

    2025

  3. [3]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y. Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  4. [4]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V. Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Gr00t n1: An open foundation model for generalist humanoid robots,

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

    Pith/arXiv arXiv 2025

  6. [6]

    Rt-2: Vision- language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt-2: Vision- language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    2023

  7. [7]

    Openvla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Bal- akrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. San- ketiet al., “Openvla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Simplevla-rl: Scal- ing vla training via reinforcement learning,

    H. Li, Y. Zuo, J. Yu, Y. Zhang, Z. Yang, K. Zhang, X. Zhu, Y. Zhang, T. Chen, G. Cuiet al., “Simplevla-rl: Scal- ing vla training via reinforcement learning,”arXiv preprint arXiv:2509.09674, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Mtil: Encoding full history with mamba for temporal imitation learning,

    Y. Zhou, Y. Lin, F. Peng, J. Chen, K. Huang, H. Yang, and Z. Yin, “Mtil: Encoding full history with mamba for temporal imitation learning,”IEEE Robotics and Automation Letters, 2025

    2025

  10. [10]

    Memoryvla: Perceptual- cognitive memory in vision-language-action models for robotic manipulation,

    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 manipulation,”arXiv preprint arXiv:2508.19236, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Robomme: Benchmarking and understanding memory for robotic generalist policies,

    Y. Dai, H. Fu, J. Lee, Y. Liu, H. Zhang, J. Yang, C. Finn, N. Fazeli, and J. Chai, “Robomme: Benchmarking and understanding memory for robotic generalist policies,” 2026. [Online]. Available: https://arxiv.org/abs/2603.04639

    Pith/arXiv arXiv 2026

  12. [12]

    History- aware visuomotor policy learning via point tracking,

    J. Chen, H. Fang, C. Wang, S. Wang, and C. Lu, “History- aware visuomotor policy learning via point tracking,”arXiv preprint arXiv:2509.17141, 2025

    arXiv 2025

  13. [13]

    Hamlet: Switch your vision-language-action model into a history-aware policy,

    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

    Pith/arXiv arXiv 2025

  14. [14]

    Diff-control: A stateful diffusion-based policy for imitation learning,

    X. Liu, Y. Zhou, F. Weigend, S. Sonawani, S. Ikemoto, and H. B. Amor, “Diff-control: A stateful diffusion-based policy for imitation learning,” in2024 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 7453–7460

    2024

  15. [15]

    Resolving state ambiguity in robot manipulation via adaptive working memory recoding,

    Q. Hu, Z. Qiu, Z. Xu, K. Zhang, X. Bu, Z. Sun, B. Zhang, J. Zhao, Z. Gan, and W. Ding, “Resolving state ambiguity in robot manipulation via adaptive working memory recoding,” arXiv preprint arXiv:2512.24638, 2025

    arXiv 2025

  16. [16]

    Echovla: Robotic vision-language-action model with synergistic declarative memory for mobile manipulation,

    M. Lin, X. Liang, B. Lin, L. Jingzhi, Z. Jiao, K. Li, Y. Ma, Y. Liu, S. Zhao, Y. Zhuang, and X. Liang, “Echovla: Robotic vision-language-action model with synergistic declarative memory for mobile manipulation,” 2025. [Online]. Available: https://arxiv.org/abs/2511.18112

    arXiv 2025

  17. [17]

    Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation,

    H. Fang, M. Grotz, W. Pumacay, Y. R. Wang, D. Fox, R. Kr- ishna, and J. Duan, “Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation,” arXiv preprint arXiv:2501.18564, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Cronusvla: Transferring latent motion across time for multi-frame prediction in ma- nipulation,

    H. Li, S. Yang, Y. Chen, Y. Tian, X. Yang, X. Chen, H. Wang, T. Wang, F. Zhao, D. Linet al., “Cronusvla: Transferring latent motion across time for multi-frame prediction in ma- nipulation,”arXiv preprint arXiv:2506.19816, 2025

    arXiv 2025

  19. [19]

    Human memory: A proposed system and its control processes,

    R. C. Atkinson and R. M. Shiffrin, “Human memory: A proposed system and its control processes,” inPsychology of learning and motivation. Elsevier, 1968, vol. 2, pp. 89–195

    1968

  20. [20]

    Episodic and semantic memory,

    E. Tulvinget al., “Episodic and semantic memory,”Organi- zation of memory, vol. 1, no. 381-403, p. 1, 1972

    1972

  21. [21]

    Working memory: An overview,

    A. Baddeley, “Working memory: An overview,”Working mem- ory and education, pp. 1–31, 2006

    2006

  22. [22]

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation,

    T. Chen, Z. Chen, B. Chen, Z. Cai, Y. Liu, Z. Li, Q. Liang, X. Lin, Y. Ge, Z. Guet al., “Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation,”arXiv preprint arXiv:2506.18088, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Rmbench: Memory- dependent robotic manipulation benchmark with insights into policy design,

    T. Chen, Y. Wang, M. Li, Y. Qin, H. Shi, Z. Li, Y. Hu, Y. Zhang, K. Wang, Y. Chenet al., “Rmbench: Memory- dependent robotic manipulation benchmark with insights into policy design,”arXiv preprint arXiv:2603.01229, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    A survey on diffusion policy for robotic manipulation: Tax- onomy, analysis, and future directions,

    M. Song, X. Deng, Z. Zhou, J. Wei, W. Guan, and L. Nie, “A survey on diffusion policy for robotic manipulation: Tax- onomy, analysis, and future directions,”Authorea Preprints, 2025

    2025

  25. [25]

    Foundation models in robotics: Applications, challenges, and the future,

    R. Firoozi, J. Tucker, S. Tian, A. Majumdar, J. Sun, W. Liu, Y. Zhu, S. Song, A. Kapoor, K. Hausmanet al., “Foundation models in robotics: Applications, challenges, and the future,” The International Journal of Robotics Research, vol. 44, no. 5, pp. 701–739, 2025

    2025

  26. [26]

    A review of learning- based dynamics models for robotic manipulation,

    B. Ai, S. Tian, H. Shi, Y. Wang, T. Pfaff, C. Tan, H. I. Christensen, H. Su, J. Wu, and Y. Li, “A review of learning- based dynamics models for robotic manipulation,”Science Robotics, vol. 10, no. 106, p. eadt1497, 2025

    2025

  27. [27]

    Rdt-1b: a diffusion foundation model for bimanual manipulation,

    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

  28. [28]

    Mo- tion before action: Diffusing object motion as manipulation condition,

    Y. Su, X. Zhan, H. Fang, Y.-L. Li, C. Lu, and L. Yang, “Mo- tion before action: Diffusing object motion as manipulation condition,”IEEE Robotics and Automation Letters, 2025

    2025

  29. [29]

    Cot-vla: Visual chain-of- thought reasoning for vision-language-action models,

    Q. Zhao, Y. Lu, M. J. Kim, Z. Fu, Z. Zhang, Y. Wu, Z. Li, Q. Ma, S. Han, C. Finnet al., “Cot-vla: Visual chain-of- thought reasoning for vision-language-action models,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1702–1713

    2025

  30. [30]

    Wmpo: World model-based policy optimization for vision-language- action models,

    F. Zhu, Z. Yan, Z. Hong, Q. Shou, X. Ma, and S. Guo, “Wmpo: World model-based policy optimization for vision-language- action models,”arXiv preprint arXiv:2511.09515, 2025

    arXiv 2025

  31. [31]

    Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,

    T. Dao and A. Gu, “Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,”arXiv preprint arXiv:2405.21060, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Rnns are not transform- ers (yet): The key bottleneck on in-context retrieval,

    K. Wen, X. Dang, and K. Lyu, “Rnns are not transform- ers (yet): The key bottleneck on in-context retrieval,”arXiv preprint arXiv:2402.18510, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Dinov2: Learning robust visual features without super- vision,

    M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec et al., “Dinov2: Learning robust visual features without super- vision,”arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  34. [34]

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

    Y. Zhong, X. Huang, R. Li, C. Zhang, Z. Chen, T. Guan, F. Zeng, K. N. Lui, Y. Ye, Y. Lianget al., “Dexgraspvla: A vision-language-action framework towards general dexterous grasping,”arXiv preprint arXiv:2502.20900, 2025

    arXiv 2025

  35. [35]

    U-net: Convo- lutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convo- lutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    2015

  36. [36]

    Denoising diffusion proba- bilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion proba- bilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  37. [37]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  38. [38]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,

    Y. Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,” inProceedings of Robotics: Science and Systems (RSS), 2024

    2024