pith. sign in

arxiv: 2606.17463 · v1 · pith:DZBQKWNUnew · submitted 2026-06-16 · 💻 cs.CV · cs.RO

WeaveLA: Event Driven Cross-Subtask Latent Memory Weaving for Repetitive Robot Manipulation

Shoujing Zhu , Zhenyang Liu , Fungmiu Wang , Jiafeng Wang , Bo Yue , Guiliang Liu , Simo Wu , Xiangyang Xue
show 1 more author
Taiping Zeng
This is my paper

Pith reviewed 2026-06-27 02:01 UTC · model grok-4.3

classification 💻 cs.CV cs.RO
keywords vision-language-actionrobot manipulationlatent memorysub-task hand-offevent-drivenrepetitive tasksVLA policies
0
0 comments X

The pith

WeaveLA adds an event-triggered latent memory channel to frozen VLA policies that compresses each completed sub-task segment into tokens and routes them directly into the next sub-task's action expert.

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

The paper sets out to establish that VLA policies can overcome brittleness at sub-task boundaries in repetitive manipulation by adding a lightweight cross-subtask channel that activates only at sub-goal completion events. It does so by compressing the just-finished segment into latent tokens through query-driven attention pooling and injecting those tokens into the action-generation path of the following sub-task, all without retraining the base model or altering its short-window interface. A sympathetic reader would care because this targets exactly the structural gap where short-window VLAs currently fail while leaving single-execution performance untouched. Stratified tests on RoboMME with a π0.5 backbone show the improvement appears only on repetition-heavy slices whose causal structure requires prior-subtask information.

Core claim

WeaveLA is a cross-subtask memory interface that, on top of a frozen VLA backbone, compresses each completed segment into latent tokens via query-driven attention pooling and routes them directly into the action-generation path of the next sub-task. This event-triggered, action-side design preserves the base policy's short-window interface while adding a lightweight cross-subtask channel. Through stratified evaluation on RoboMME with a π0.5 backbone, success on the hardest repetition slice (SwingXtimes, N=3) rises from 0% to 47.8%, while single-execution episodes remain unchanged. Per-episode paired analysis confirms the gains are confined to tasks whose causal structure requires cross-subta

What carries the argument

The event-triggered cross-subtask memory interface that uses query-driven attention pooling to produce latent tokens for hand-off into the action expert of the subsequent sub-task.

If this is right

  • Success rates rise on the hardest repetition tasks while single-execution tasks stay the same.
  • Gains appear only on tasks whose causal structure requires cross-subtask information.
  • The base policy's short-window interface and overall behavior on non-repetitive episodes are preserved.
  • The added channel remains lightweight because it activates only at sub-goal events rather than every frame.

Where Pith is reading between the lines

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

  • The same event-triggered hand-off pattern could be tested on sequential tasks outside robot manipulation where prior-step information must reach later actions.
  • Replacing the query-driven pooling with other compression methods would isolate whether the attention mechanism itself is load-bearing.
  • Placing the memory tokens on the observation side instead of the action side could be compared to measure the effect of routing choice.

Load-bearing premise

The sub-goal completion event is the natural temporal unit for cross-subtask memory hand-off and the latent tokens it produces are useful when fed into the action expert of the next sub-task.

What would settle it

Adding the memory-weaving channel would leave success rates on repetition-heavy tasks such as SwingXtimes N=3 unchanged or would degrade performance on single-execution episodes.

Figures

Figures reproduced from arXiv: 2606.17463 by Bo Yue, Fungmiu Wang, Guiliang Liu, Jiafeng Wang, Shoujing Zhu, Simo Wu, Taiping Zeng, Xiangyang Xue, Zhenyang Liu.

Figure 1
Figure 1. Figure 1: Sub-goal-event trigger storyboard. For an episode that decomposes into three sub-goals, the Weaver fires exactly at each completion event, producing one fresh memory latent that conditions the action expert during the next sub-task. The trigger is event-driven, not per-frame [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: WeaveLA framework. (A) A frozen VLM/decoder backbone extracts perceptual features from the previous sub-task’s observations, language instruction, and state history. (B) The Query￾driven Memory Weaver concatenates visual, text, and state features into mixed hidden states, then compresses them into 8 latent memory tokens via a single attention-pooling step driven by 8 learnable query latents. (C) During act… view at source ↗
Figure 3
Figure 3. Figure 3: Aggregate π0.5+AP results on the six trained tasks (success ↑, %, 50 episodes). Improve￾ments concentrate on STOPC/SWINGX at both training scales. “Weaver off (no-weaving baseline)” uses our framework with the memory-weaving pathway disabled, sharing training data, optimiser with the corresponding Weaver-on run. SWINGXTIMES) and 16-task (all RoboMME tasks); all four π0.5+AP runs (Weaver-on/off × 6- task/16… view at source ↗
Figure 4
Figure 4. Figure 4: Repetition-stratified success on SWINGXTIMES and PICKXTIMES (success ↑, %; 6-task training; N = number of sub-goals encoded in the instruction). Annotations in the figure are rounded to integer percentages; the underlying value at N=3 is 47.8% (rounded to 48%). Both conditions reach ∼ 100% at N=1; the gap opens at N≥2. STOPCUBE aggregate Weaver-on/off levels are overlaid as horizontal references. concentra… view at source ↗
Figure 5
Figure 5. Figure 5: STOPCUBE as a what/where/when ablation across the RoboMME baseline (success ↑, %, 50 episodes). WeaveLA (red) is plotted against published MME-VLA variants and prior-method baselines. Each panel reorganises the same set of methods by one axis of the memory design [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-frame comparison of Weaver-on vs. Weaver-off on two timing-critical tasks under identical initial conditions. Top two rows: SWINGXTIMES — pick up the green cube, swing it between the right and left targets three times, then press the stop button. Bottom two rows: STOPCUBE — press the button at the exact moment the oscillating cube reaches the target for the third time. Without Weaver (rows 1 and 3), th… view at source ↗
Figure 7
Figure 7. Figure 7: weaver_latent_norm during training, 16-task scale. The two staged-training runs (stage2_16task, stage2_16task_from_stage1) keep the latent norm stable at ∼ 46 throughout training. The merged-stage run (16task_hybrid) shows monotonic decay starting around step 8k, ending near ∼ 38 at step 16k. Downstream evaluation success on the merged-stage run collapsed to near zero, despite stable action and alignment l… view at source ↗
Figure 8
Figure 8. Figure 8: Real-robot PICKXTIMES platform. (a) Side perspective view, showing the PiPER arm, the third-person camera, the wrist camera, and the manipulation objects on the workspace. (b) Top-down view, showing the Intel RealSense monitoring camera, the gripper, and the layout of the three coloured blocks and the white plate [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: PICKXTIMES task example. In this instance, the goal is to pick up the blue cube and place it on the target, repeating this pick-and-place action three times, then return to default to stop. Frames are sampled at sub-goal completion events from a successful N=3 rollout, illustrating the three pick–place repetitions and the return-to-default termination. (e.g. [30, 54, 88] for a 3-sub-goal episode) and consu… view at source ↗
Figure 10
Figure 10. Figure 10: Web-based sub-goal boundary annotator. The four camera streams are displayed synchronously; the annotator scrubs through frames using the speed buttons (0.5x–10x) or per-frame controls and presses Mark / Unmark at each sub-goal completion event. The current frame index and the list of saved frame indices for the active episode are shown below the playback controls; an episode is finalised once all sub-goa… view at source ↗
read the original abstract

Vision-Language-Action (VLA) policies have achieved remarkable single-step manipulation, yet they remain brittle precisely where each stage depends on what was just completed. The core issue is structural: short-window VLAs lack an explicit channel for rouxting information across sub-task boundaries, and existing memory-augmented variants either write at every frame, retrieve from demonstration-time stages, or fire at sub-goal events without performing an explicit sub-task-to-sub-task hand-off into the action expert. We identify the sub-goal completion event as the natural temporal unit for cross-subtask memory hand-off, and present WeaveLA (Weave Latent memory for Vision-Language-Action policies), a cross-subtask memory interface that, on top of a frozen VLA backbone, compresses each completed segment into latent tokens via query-driven attention pooling and routes them directly into the action-generation path of the next sub-task. This event-triggered, action-side design preserves the base policy's short-window interface while adding a lightweight cross-subtask channel. Through stratified evaluation on RoboMME with a $\pi_{0.5}$ backbone, WeaveLA's gains land exactly where the channel is needed: on the hardest repetition slice (SwingXtimes, $N{=}3$), success rises from $0\%$ to $47.8\%$, while single-execution episodes remain unchanged. Per-episode paired analysis confirms the gains are confined to tasks whose causal structure requires cross-subtask information.

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

3 major / 2 minor

Summary. The paper introduces WeaveLA, an event-driven cross-subtask latent memory interface for frozen VLA backbones. Sub-goal completion events trigger query-driven attention pooling to compress completed segments into latent tokens that are routed directly into the action expert of the subsequent sub-task. On the RoboMME benchmark with a π0.5 backbone, the method reports large gains precisely on the hardest repetition slice (SwingXtimes, N=3: 0% → 47.8% success) while leaving single-execution episodes unchanged; per-episode analysis is used to argue that improvements are confined to tasks whose causal structure requires cross-subtask information.

Significance. If the central result holds, the work supplies a lightweight, action-side memory channel that preserves the short-window interface of existing VLAs while addressing brittleness at sub-task boundaries. The stratified evaluation design and the claim that gains appear only where cross-subtask state is required constitute a falsifiable prediction that strengthens the contribution. The approach is compatible with frozen backbones, which is practically relevant.

major comments (3)
  1. [Experiments section] Experiments section (stratified evaluation on RoboMME): the headline 0%→47.8% lift on SwingXtimes (N=3) is attributed to the event-triggered hand-off of query-driven pooled latents, yet no ablation keeps the sub-goal event trigger and routing fixed while replacing the pooled tokens with uninformative content (zero vectors, random vectors, or tokens from a different operator). Without this control, the observed gain cannot be credited specifically to the cross-subtask channel rather than to the mere presence of an event-driven interface.
  2. [Methods] Methods, query-driven attention pooling paragraph: the description states that each completed segment is compressed into latent tokens via query-driven attention pooling and routed into the action-generation path, but does not specify how the queries are constructed (learned parameters, fixed templates, or derived from the VLA's own embeddings) or whether the pooling operator is trained. This detail is load-bearing for the claim that the tokens carry usable cross-subtask state into the frozen π0.5 action expert.
  3. [Experiments section] Evaluation protocol: success rates are reported without error bars, number of evaluation seeds, or statistical tests. Given that the central claim rests on a large relative improvement on a single slice (SwingXtimes N=3), the absence of these quantities makes it impossible to assess whether the reported 47.8% is robust or could be explained by evaluation variance.
minor comments (2)
  1. [Abstract] The abstract and introduction use the non-standard spelling 'rouxting'; this should be corrected to 'routing'.
  2. [Throughout] Notation for the backbone is written as both π0.5 and π_{0.5}; adopt a single consistent form throughout.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and commit to revisions that directly strengthen the experimental controls and reporting.

read point-by-point responses

  1. Referee: [Experiments section] Experiments section (stratified evaluation on RoboMME): the headline 0%→47.8% lift on SwingXtimes (N=3) is attributed to the event-triggered hand-off of query-driven pooled latents, yet no ablation keeps the sub-goal event trigger and routing fixed while replacing the pooled tokens with uninformative content (zero vectors, random vectors, or tokens from a different operator). Without this control, the observed gain cannot be credited specifically to the cross-subtask channel rather than to the mere presence of an event-driven interface.

    Authors: We agree that the current experiments lack a control that isolates the informational content of the pooled latents. We will add an ablation that retains the sub-goal event trigger and routing mechanism but substitutes the pooled tokens with zero vectors (and separately with random vectors) and report the resulting success rates on the SwingXtimes (N=3) slice. revision: yes

  2. Referee: [Methods] Methods, query-driven attention pooling paragraph: the description states that each completed segment is compressed into latent tokens via query-driven attention pooling and routed into the action-generation path, but does not specify how the queries are constructed (learned parameters, fixed templates, or derived from the VLA's own embeddings) or whether the pooling operator is trained. This detail is load-bearing for the claim that the tokens carry usable cross-subtask state into the frozen π0.5 action expert.

    Authors: The queries are implemented as learned parameters and the pooling operator is trained end-to-end as part of the interface. We will revise the methods paragraph to explicitly state the query construction (learned parameters initialized from VLA embeddings) and confirm that the pooling operator is trained while the backbone remains frozen. revision: yes

  3. Referee: [Experiments section] Evaluation protocol: success rates are reported without error bars, number of evaluation seeds, or statistical tests. Given that the central claim rests on a large relative improvement on a single slice (SwingXtimes N=3), the absence of these quantities makes it impossible to assess whether the reported 47.8% is robust or could be explained by evaluation variance.

    Authors: We acknowledge that variability measures are required for the central claim. We will re-run the key evaluations over multiple random seeds, add error bars to all reported success rates, and include statistical tests comparing WeaveLA against the baseline on the repetition slices. revision: yes

Circularity Check

0 steps flagged

No significant circularity; additive interface with external empirical validation

full rationale

The paper describes WeaveLA as a new event-triggered memory interface added to a frozen π0.5 VLA backbone, using query-driven attention pooling to produce latent tokens routed to subsequent sub-tasks. Claims rest on stratified benchmark results (e.g., 0% to 47.8% on SwingXtimes N=3) rather than any derivation, equation, or self-citation that reduces the reported gains to quantities defined by the method itself. No load-bearing steps reduce by construction to fitted inputs or author prior work; the architecture is presented as an independent additive channel whose utility is assessed on external tasks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method description mentions only standard attention pooling and frozen backbone usage.

pith-pipeline@v0.9.1-grok · 5829 in / 1230 out tokens · 43224 ms · 2026-06-27T02:01:32.342020+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

53 extracted references · 6 linked inside Pith

  1. [1]

    J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization, 2016.arXiv preprint arXiv:1607.06450

    Pith/arXiv arXiv 2016

  2. [2]

    Bacon, J

    P.-L. Bacon, J. Harb, and D. Precup. The option-critic architecture. InAAAI, 2017

    2017

  3. [3]

    Beyer, A

    L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, et al. PaliGemma: A versatile 3B VLM for transfer, 2024.arXiv preprint arXiv:2407.07726

    Pith/arXiv arXiv 2024

  4. [4]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, et al. π0: A vision-language-action flow model for general robot control, 2024

    2024

  5. [5]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control, 2023

    2023

  6. [6]

    Bulatov, Y

    A. Bulatov, Y . Kuratov, and M. S. Burtsev. Recurrent memory transformer, 2022

    2022

  7. [7]

    Sobol Mark, J

    M. Sobol Mark, J. Liang, M. Attarian, C. Fu, D. Dwibedi, D. Shah, and A. Kumar. BPP: Long-context robot imitation learning by focusing on key history frames, 2026.arXiv preprint arXiv:2602.15010

    arXiv 2026

  8. [8]

    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.arXiv preprint arXiv:2603.04639, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Dwibedi, Y

    D. Dwibedi, Y . Aytar, J. Tompson, P. Sermanet, and A. Zisserman. Counting out time: Class-agnostic video repetition counting in the wild, 2020

    2020

  10. [10]

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

    2025

  11. [11]

    Gemma: Open models based on Gemini research and technology, 2024a.arXiv preprint arXiv:2403.08295

    Gemma Team. Gemma: Open models based on Gemini research and technology, 2024a.arXiv preprint arXiv:2403.08295

    Pith/arXiv arXiv

  12. [12]

    Gemma 2: Improving open language models at a practical size, 2024b.arXiv preprint arXiv:2408.00118

    Gemma Team. Gemma 2: Improving open language models at a practical size, 2024b.arXiv preprint arXiv:2408.00118

    Pith/arXiv arXiv

  13. [13]

    S. Han, B. Qiu, Y . Liao, S. Huang, C. Gao, S. Yan, and S. Liu. RoboCerebra: A large-scale benchmark for long-horizon robotic manipulation evaluation, 2025

    2025

  14. [14]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, et al. LoRA: Low-rank adaptation of large language models, 2022

    2022

  15. [15]

    H. Hu, S. Dong, Y . Zhao, D. Lian, Z. Li, and S. Gao. TransRAC: Encoding multi-scale temporal correlation with transformers for repetitive action counting, 2022

    2022

  16. [16]

    H. Jang, S. Yu, H. Kwon, H. Jeon, Y . Seo, and J. Shin. ContextVLA: Vision-language-action model with amortized multi-frame context, 2025

    2025

  17. [17]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, et al. OpenVLA: An open-source vision-language-action model, 2024

    2024

  18. [18]

    T. Kipf, Y . Li, H. Dai, V . Zambaldi, A. Sanchez-Gonzalez, E. Grefenstette, P. Kohli, and P. Battaglia. CompILE: Compositional imitation learning and execution. InICML, 2019

    2019

  19. [19]

    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, 2025.arXiv preprint arXiv:2510.00695

    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, 2026.arXiv preprint arXiv:2603.12942. 10

    arXiv 2026

  21. [21]

    J. Li, D. Li, S. Savarese, and S. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023

    2023

  22. [22]

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

    arXiv 2025

  23. [23]

    Lipman, R

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

    2022

  24. [24]

    Y . Fan, P. Ding, S. Bai, X. Tong, Y . Zhu, et al. Long-VLA: Unleashing long-horizon capability of vision-language-action model for robot manipulation. InCoRL, 2025

    2025

  25. [25]

    Bjorck, F

    NVIDIA, J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, et al. GR00T N1: An open foundation model for generalist humanoid robots, 2025

    2025

  26. [26]

    GR00T N1.6: An improved open foundation model for generalist humanoid robots

    NVIDIA GEAR Team. GR00T N1.6: An improved open foundation model for generalist humanoid robots. https://research.nvidia.com/labs/gear/gr00t-n1_6, 2025

    2025

  27. [27]

    Paiss, A

    R. Paiss, A. Ephrat, O. Tov, S. Zada, I. Mosseri, M. Irani, and T. Dekel. Teaching CLIP to count to ten, 2023

    2023

  28. [28]

    Z. Liu, Y . Wang, S. Zheng, T. Pan, L. Liang, Y . Fu, and X. Xue. ReasonGrounder: LVLM-guided hierarchical feature splatting for open-vocabulary 3D visual grounding and reasoning. InCVPR, 2025a

  29. [29]

    Black, N

    Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, et al. π0.5: A vision-language-action model with open-world generalization, 2025

    2025

  30. [30]

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu, et al. MemoryVLA: Perceptual-cognitive memory in vision- language-action models for robotic manipulation, 2025

    2025

  31. [31]

    Sinha, A

    S. Sinha, A. Stergiou, and D. Damen. Every shot counts: Using exemplars for repetition counting in videos. InACCV, 2024

    2024

  32. [32]

    Sridhar, J

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

    arXiv 2025

  33. [33]

    J. Su, A. Murtadha, Y . Lu, S. Pan, Wen. Bo, and Y . Liu. RoFormer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  34. [34]

    Z. Liu, S. Zheng, S. Chen, C. Zhao, L. Liang, X. Xue, and Y . Fu. A neural representation framework with LLM-driven spatial reasoning for open-vocabulary 3D visual grounding. InACM MM, 2025b

  35. [35]

    R. S. Sutton, D. Precup, and S. Singh. Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning.Artificial Intelligence, 112(1–2):181–211, 1999

    1999

  36. [36]

    Torne, K

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

    2026

  37. [37]

    Toro Icarte, T

    R. Toro Icarte, T. Q. Klassen, R. Valenzano, and S. A. McIlraith. Using reward machines for high-level task specification and decomposition in reinforcement learning. InICML, 2018

    2018

  38. [38]

    Vaezipoor, A

    P. Vaezipoor, A. C. Li, R. Toro Icarte, and S. A. McIlraith. LTL2Action: Generalizing LTL instructions for multi-task RL. InICML, 2021

    2021

  39. [39]

    Z. Liu, Y . Wang, K. Wang, L. Liang, X. Xue, and Y . Fu. Spatial-temporal aware visuomotor diffusion policy learning. InICCV, 2025c

  40. [40]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. InNeurIPS, 2017

    2017

  41. [41]

    X. Guo, Z. Huang, Z. Shi, Z. Song, and J. Zhang. Your vision-language model can’t even count to 20: Exposing the failures of VLMs in compositional counting, 2025.arXiv preprint arXiv:2510.04401

    arXiv 2025

  42. [42]

    Zhang, Y

    J. Zhang, Y . Guo, X. Chen, Y .-J. Wang, Y . Hu, C. Shi, and J. Chen. HiRT: Enhancing robotic control with hierarchical robot transformers. InCoRL, 2024

    2024

  43. [43]

    Zheng, Y

    R. Zheng, Y . Liang, S. Huang, J. Gao, H. Daumé III, A. Kolobov, F. Huang, and J. Yang. TraceVLA: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies, 2024. 11

    2024

  44. [44]

    Z. Liu, Y . Gu, S. Zheng, Y . Fu, X. Xue and Y . Jiang. TriVLA: A triple-system-based unified vision- language-action model for general robot control, 2025d.arXiv preprint arXiv:2507.01424

    arXiv

  45. [45]

    Z. Liu, Y . Gu, Y . Wang, X. Xue, and Y . Fu. ActiveVLA: Injecting active perception into vision-language- action models for precise 3D robotic manipulation, 2026.arXiv preprint arXiv:2601.08325

    arXiv 2026

  46. [46]

    Zhang, M

    G. Zhang, M. Fu, and S. Yan. MemGen: Weaving generative latent memory for self-evolving agents, 2025. arXiv preprint arXiv:2509.24704

    arXiv 2025

  47. [47]

    Zhang and R

    B. Zhang and R. Sennrich. Root mean square layer normalization. InNeurIPS, 2019

    2019

  48. [48]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InRobotics: Science and Systems (RSS), 2023a

  49. [49]

    Y . Chen, W. Tan, L. Zhu, F. Li, J. Li, G. Yang, and H. T. Shen. Non-Markovian long-horizon robot manipulation via keyframe chaining, 2026.arXiv preprint arXiv:2603.01465

    arXiv 2026

  50. [50]

    H. Wu, T. Chen, J. Wang, X. Li, and L. Fang. StreamVLA: Breaking the reason-act cycle via completion- state gating, 2026.arXiv preprint arXiv:2602.01100

    arXiv 2026

  51. [51]

    R. Yang, Z. An, L. Zhou, and Y . Feng. SeqVLA: Sequential task execution for long-horizon manipulation with completion-aware vision-language-action model, 2025.arXiv preprint arXiv:2509.14138

    arXiv 2025

  52. [52]

    W. Wan, Y . Zhu, R. Shah, and Y . Zhu. LOTUS: Continual imitation learning for robot manipulation through unsupervised skill discovery. InICRA, 2024

    2024

  53. [53]

    Weaver off (no-weaving baseline)

    Y . Zhu, P. Stone, and Y . Zhu. Bottom-up skill discovery from unsegmented demonstrations for long-horizon robot manipulation.IEEE Robotics and Automation Letters, 2022. A Technical Appendices and Supplementary Material A.1 Full 16-Task RoboMME Breakdown Table 2: Full RoboMME 16-task success rates (↑, %, 50 episodes per task) for the four π0.5+AP con- fig...

    2022