Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Caching past robot actions cuts VLA inference latency by up to 34×

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 · glm-5.2

2026-07-08 08:03 UTC pith:3DHMPXNN

load-bearing objection Training-free action caching for flow-based VLA models: real within-task speedups, but headline numbers are oversold and cross-task generalization is weak. the 4 major comments →

arxiv 2607.06370 v1 pith:3DHMPXNN submitted 2026-07-07 cs.RO cs.CVcs.LG

Training-Free Acceleration for Vision-Language-Action Models with Action Caching and Refinement

Ryuji Oi , Hikari Otsuka , Kosuke Matsushima , Yuki Ichikawa , Masato Motomura , Tatsuya Kaneko , Daichi Fujiki This is my paper
classification cs.RO cs.CVcs.LG
keywords vision-language-action modelsflow matchingrobotic manipulationinference accelerationaction cachingwarm startingdenoisingtraining-free methods
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.

This paper claims that the iterative denoising process used by flow-based Vision-Language-Action (VLA) robot policies can be dramatically shortened by reusing action chunks from past successful generations when the current visual and linguistic context is sufficiently similar to a cached one. The authors introduce ActionCache, a training-free, plug-and-play external memory that stores intermediate action chunks alongside compact keys derived from VLM output embeddings compressed via sparse ternary random projection. When a new control step produces a key whose cosine similarity to a cached entry exceeds a threshold, the cached action chunk is used to warm-start the flow matching process instead of starting from Gaussian noise, reducing the number of denoising steps required—sometimes to zero. The paper reports acceleration of up to 11.75× for π0.5 and 34.43× for GR00T-N1.6 while maintaining task success rates comparable to full-step generation in both simulation and real-world pick-and-place experiments.

Core claim

The central finding is that robot manipulation tasks contain substantial redundancy in their visual and linguistic contexts, and that this redundancy induces similar conditional flow transport paths in the action generation process. By compressing VLM output embeddings into 500-dimensional sparse ternary random projection keys and using cosine similarity to retrieve cached intermediate action chunks, ActionCache can identify when a past action generation is close enough to the current target to serve as a warm start. This reduces or eliminates the denoising iterations that dominate inference latency, with the hit threshold Thit serving as a tunable knob controlling the quality-latency trade-

What carries the argument

The mechanism is a retrieve-and-refine loop: (1) VLM output embeddings are projected into a 500-dim key via a fixed sparse ternary random matrix; (2) cosine similarity against cached keys determines a hit or miss via threshold Thit; (3) on a hit, the cached intermediate action chunk (stored at flow time τ = (N − Nhit)Δτ) initializes flow matching with Nhit or fewer denoising steps; (4) on a miss, full denoising from Gaussian noise proceeds as normal; (5) only actions from successful episodes are committed to the cache, managed by LFU replacement at limited capacity.

Load-bearing premise

The method assumes that cosine similarity in a 500-dimensional sparse random projection of VLM output embeddings is a reliable proxy for whether a cached intermediate action chunk will be close enough to the current target action to serve as a useful warm start. The need to tune the similarity threshold differently for each model and task setting suggests this proxy is not universally calibrated.

What would settle it

A scenario in which two visually and linguistically similar contexts require substantially different action chunks—due to unobserved state differences, multimodal action distributions, or fine-grained spatial variations—would cause the cache to retrieve a misleading warm start, potentially degrading task success below that of simple noise-initialized generation. This would be most likely in tasks with high sensitivity to small visual differences or in environments with greater diversity than those tested.

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

If this is right

  • If action-level caching proves robust across more visually diverse environments, it could become a standard inference component for deployed robot policies, making real-time closed-loop control feasible on lower-power hardware.
  • The cross-task hit rates observed in early episode stages suggest that shared primitive motions (approaching, reaching) could be pre-computed and distributed as cached action libraries for new tasks.
  • The finding that LFU outperforms LRU under limited cache capacity implies that frequently-reused actions are more valuable than recently-used ones, which could inform the design of action libraries that prioritize common robust primitives over temporal locality.
  • The zero-NFE mode (directly executing cached actions without any denoising) suggests that for sufficiently similar contexts, the flow matching process is entirely redundant, raising the question of whether some action generation steps are unnecessary for specific task phases.
  • Combining ActionCache with VLM backbone acceleration methods could yield compounding speedups, since the two approaches target different components of the inference pipeline.

Where Pith is reading between the lines

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

  • The per-model and per-task tuning of Thit (0.85 for π0.5, 0.65 for GR00T-N1.6, 0.925 for cross-task) suggests the compressed key's discriminative power is model- and domain-dependent; a learned or adaptive key projection might reduce this tuning burden and extend applicability.
  • The drop in cache hit rate during grasping phases (shown in real-world experiments) implies that the method's acceleration benefit is uneven across task phases, and that overall speedup factors may overstate gains for contact-rich or highly variable manipulation steps.
  • If the approach were extended to environments with higher visual diversity or multimodal action distributions, the hit rate may drop substantially, potentially making the method less effective than simple NFE reduction in those settings.

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 / 6 minor

Summary. This paper proposes ActionCache, a training-free, plug-and-play external cache that accelerates flow-based VLA models (π0.5 and GR00T-N1.6) by storing intermediate action chunks indexed by compact multimodal keys derived from VLM output embeddings via sparse ternary random projection. On a cache hit, the retrieved action chunk warm-starts the flow matching process, reducing the number of denoising steps (NFE). The method includes a conservative fallback to full-step generation on cache misses, and supports multiple replacement policies (LRU, LFU, FIFO). Experiments span VLABench and LIBERO simulations, a real-world pick-and-place task, and ablations on cache key source, replacement policy, retrieval quality, and cross-task reuse.

Significance. The paper addresses a practical bottleneck (action head latency) in flow-based VLA models with a genuinely training-free approach that requires no learned components—the sparse ternary projection matrix is fixed and parameter-free (§3.2, Eq. 2). The ablation studies are thorough: Table 3 compares cache key sources, Figure 5 characterizes the retrieval-quality–success-rate relationship, Figure 6 compares replacement policies, and Table 2 provides a cross-task reuse experiment. The real-world deployment (Table 4, Figure 7) with phase-specific hit rate analysis adds practical value. The method is model-agnostic and could complement existing VLM-level acceleration methods. These are real strengths.

major comments (4)
  1. Abstract and §4.2: The claim that ActionCache 'maintains high task success rates' is not supported by the headline speedup numbers. For GR00T-N1.6 at NFE=0 (the 34.43× speedup), the success rate drops from 34.0% to 22.3%—a 34% relative degradation (Table 1). For π0.5 at NFE=0 (the 11.75× speedup), the drop is 38.8%→32.9%, a 15% relative degradation. While moderate speedups (e.g., 3.48× at NFE=2 for π0.5 with only -7.8% absolute drop) are more defensible, the abstract's framing of 'maintaining high task success rates' alongside the headline numbers is misleading. The authors should either recalibrate the abstract to acknowledge the trade-off at aggressive speedups or foreground the moderate-speedup results where success rates are genuinely maintained.
  2. Abstract and §4.3: The abstract claims the cache 'enables retrieval from similar past contexts across different episodes or even different tasks.' However, the only cross-task experiment (§4.3, Table 2) shows hit rates of 2.7% (select_painting) and 8.1% (select_toy) when the cache is populated with select_fruit. These rates mean the cache is almost never useful cross-task. While the authors note that hit rates are higher at episode start (Figure 4), the overall cross-task utility is marginal. The abstract's claim about cross-task retrieval should be tempered to reflect this evidence, or additional cross-task experiments with higher hit rates should be provided.
  3. §4.1 and Table 1: The warm-up protocol is insufficiently specified. Table 1 evaluates on 'Seen' object configurations, but the paper does not state whether warm-up episodes use the same or different object configurations than evaluation episodes. This distinction is load-bearing: if warm-up and evaluation use the same configurations, the reported hit rates may partly reflect near-duplicate retrieval rather than genuine context similarity. The authors should clarify this protocol and, ideally, report results on unseen configurations to assess how much of the speedup generalizes beyond memorized states.
  4. §4.1: The hit threshold Thit is tuned per model (0.85 for π0.5, 0.65 for GR00T-N1.6) and per setting (0.925 for cross-task in §4.3). The large gap between models (0.85 vs. 0.65) suggests the discriminative power of the compressed key varies substantially across architectures, and the need for a much higher threshold in cross-task settings suggests the key space is not well-calibrated for out-of-distribution retrieval. The paper should discuss this sensitivity more explicitly and ideally provide guidance or a protocol for setting Thit without per-model tuning.
minor comments (6)
  1. Table 1: The 'Diff' column for GR00T-N1.6 at NFE=0 shows -11.7, but 34.0→22.3 is actually -11.7 percentage points (correct), while the relative drop is ~34%. Clarifying whether 'Diff' is absolute or relative would help readers interpret the trade-off.
  2. §3.3: The warm-up procedure ('warmed up with Thit = 1 until the cache reaches maximum capacity') is described briefly. It would help to specify how many warm-up episodes are used and whether the cache is fully populated before evaluation begins.
  3. Figure 3a: The x-axis label 'NFE' could be confused with the total NFE of the base model vs. Nhit. Clarifying that the x-axis represents Nhit (NFE on cache hit) would improve readability.
  4. Table 4: The real-world evaluation uses only a single pick-and-place task. While informative, a brief discussion of whether the hit-rate patterns (high during approach/place, low during grasp) would generalize to more dexterous tasks would strengthen the discussion.
  5. Appendix B, Table 5: The LIBERO results show that the base model already achieves 89% success at NFE=1, limiting the headroom for ActionCache. The paper acknowledges this ('simple tasks can be solved with very few denoising steps'), but it would be worth noting in the main text that the method's benefit is task-difficulty-dependent.
  6. References: The citation for π0 [5] lists arXiv:2410.24164 with year 2026, which appears to be a typo for 2024.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee correctly identifies real strengths of the paper (training-free design, thorough ablations, real-world validation) and raises four major comments that are largely valid. We address each below and describe the revisions we will make.

read point-by-point responses

  1. Referee: Abstract and §4.2: The claim that ActionCache 'maintains high task success rates' is not supported by the headline speedup numbers. For GR00T-N1.6 at NFE=0 (the 34.43× speedup), the success rate drops from 34.0% to 22.3%—a 34% relative degradation. For π0.5 at NFE=0, the drop is 38.8%→32.9%, a 15% relative degradation. The abstract's framing is misleading.

    Authors: The referee is correct. The abstract's current framing pairs 'maintains high task success rates' with the headline speedup numbers (11.75× and 34.43×) in a way that overstates the success-rate preservation at those most aggressive operating points. We will revise the abstract to explicitly acknowledge the trade-off: at moderate speedups (e.g., 3.48× for π0.5 at NFE=2 with only -7.8% absolute drop, or 1.79× for GR00T-N1.6 at NFE=2 with +1.7% gain), success rates are genuinely maintained or even improved; at the most aggressive speedups, some degradation is incurred. We will also adjust §4.2 to foreground the moderate-NFE results as the primary recommended operating regime and frame the NFE=0 results as demonstrating the upper bound of achievable speedup rather than as a 'maintained' success rate. revision: yes

  2. Referee: Abstract and §4.3: The abstract claims the cache 'enables retrieval from similar past contexts across different episodes or even different tasks.' However, the only cross-task experiment shows hit rates of 2.7% and 8.1%, which means the cache is almost never useful cross-task. The abstract's claim should be tempered.

    Authors: We agree that the overall cross-task hit rates (2.7% and 8.1%) are low and that the abstract overstates the cross-task utility. We will revise the abstract to say that ActionCache 'supports retrieval from similar past contexts across different episodes, with limited but non-trivial cross-task reuse in shared task phases.' In §4.3, we will more clearly frame the cross-task experiment as a probe of potential rather than a practically deployable capability: the low overall hit rates reflect the fact that most task phases are task-specific, while the elevated hit rates at episode start (over 80%, Figure 4) show that early-phase motions (e.g., reaching toward a workspace region) are shared. We believe this is a genuine and interesting finding, but we agree the abstract should not imply broadly useful cross-task retrieval. revision: yes

  3. Referee: §4.1 and Table 1: The warm-up protocol is insufficiently specified. It is unclear whether warm-up episodes use the same or different object configurations than evaluation episodes. If they are the same, the reported hit rates may partly reflect near-duplicate retrieval rather than genuine context similarity. The authors should clarify and ideally report results on unseen configurations.

    Authors: We agree that this distinction is important and that the manuscript should have been explicit. To clarify: in the VLABench evaluation, the 'Seen' object configurations refer to the same set of object configurations used during both warm-up and evaluation. The referee is right that this means some cache hits may reflect retrieval from near-duplicate visual states rather than genuinely similar but distinct contexts. We will (1) explicitly state the warm-up protocol in §4.1, (2) acknowledge this limitation in the discussion of Table 1, and (3) add results on unseen object configurations (configurations not present during warm-up) to assess generalization. We expect hit rates to be somewhat lower on unseen configurations, particularly for task phases with high visual specificity (e.g., grasping), but we anticipate that structurally repetitive phases (approaching, placing) will still yield meaningful hits. We will report these results transparently regardless of the outcome. revision: yes

  4. Referee: §4.1: The hit threshold Thit is tuned per model (0.85 for π0.5, 0.65 for GR00T-N1.6) and per setting (0.925 for cross-task). The large gap suggests the discriminative power of the compressed key varies across architectures, and the need for a much higher threshold in cross-task settings suggests the key space is not well-calibrated for out-of-distribution retrieval. The paper should discuss this sensitivity and provide guidance for setting Thit without per-model tuning.

    Authors: This is a fair observation. The gap between 0.85 and 0.65 does reflect differences in the embedding distributions of the two architectures: π0.5's VLM output embeddings produce a more concentrated similarity distribution, requiring a higher threshold to discriminate, while GR00T-N1.6's (which include robot-state features) produce a wider distribution. The elevated 0.925 for cross-task is deliberately conservative to prevent accepting low-quality retrievals from a different task distribution. We agree this should be discussed more explicitly. In the revision, we will: (1) add a paragraph in §4.4 analyzing why thresholds differ across models, referencing the similarity distributions shown in Figure 5b; (2) provide a practical protocol: set Thit by first measuring the Top-1 similarity distribution on a small validation set and choosing a threshold at a high percentile (e.g., 90th) of the within-task hit distribution; (3) acknowledge that per-model calibration is currently required and discuss this as a limitation. We cannot honestly claim a fully automatic threshold-setting procedure at this stage, and we will say so. revision: partial

Circularity Check

0 steps flagged

No circularity found: the method is training-free, uses no learned components, and the sparse ternary random projection is fixed and parameter-free.

full rationale

The paper proposes ActionCache, a training-free caching method for flow-based VLA models. The core mechanism—storing intermediate action chunks with compact multimodal keys and retrieving them via cosine similarity—is not defined in terms of the quantities it predicts. The sparse ternary random projection matrix R (Eq. 2) is fixed and parameter-free, requiring no fitting data. The cached actions are generated by the base VLA model itself and used as initializations for refinement under the current condition, not as direct outputs (except at NFE=0, which is explicitly reported with its success-rate trade-off). The hit threshold Thit is tuned per model, but this is a hyperparameter selection, not a fitted input renamed as a prediction. No self-citation chain is load-bearing: the method is self-contained and validated against external benchmarks (VLABench, LIBERO, real-world). The cross-task experiment (Section 4.3) shows low hit rates (2.7–8.1%), which the skeptic flags as a generalization concern, but this is a correctness/evaluation-design issue, not circularity. The headline speedup numbers at NFE=0 come with explicitly reported success-rate drops (Table 1), so the paper does not hide the trade-off. No step in the derivation chain reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

ActionCache introduces no new physical entities, forces, particles, or mathematical objects. The sparse ternary random projection matrix R is a standard construction from the dimensionality reduction literature [18]. The cache itself is a data structure, not a new theoretical entity. All axioms are domain assumptions about the structure of robot manipulation tasks and the informativeness of VLM embeddings, not ad hoc postulates invented to make the derivation work.

free parameters (6)
  • Cache key dimension d = 500
    Chosen by the authors; controls the compression of VLM embeddings. Not tuned systematically in the paper.
  • Sparse projection density p = 0.01
    Fraction of non-zero entries in the random projection matrix R. Fixed by the authors without systematic justification.
  • Cache size = 3000 (default), 1000 (real-world), 50-5000 (ablation)
    Maximum number of cache entries. Default chosen by authors; ablated in Figure 5.
  • Hit threshold Thit = 0.85 (π0.5), 0.65 (GR00T-N1.6), 0.925 (cross-task)
    Per-model tuned threshold for cache hit acceptance. Different values for different models suggest it is fitted to each model's key distribution.
  • NFE on cache hit (Nhit) = 0, 1, or 2 (varies by experiment)
    Number of denoising steps after cache hit. Swept in experiments to show trade-off.
  • Action execution horizon = 10 (simulation), 50 (LIBERO)
    Number of actions executed per chunk. Set to match base model defaults.
axioms (4)
  • domain assumption VLM output embeddings encode action-relevant multimodal context that is discriminative for identifying similar conditional flow transport paths.
    Invoked in §3.2: 'these embeddings encode visual and linguistic context after backbone processing, making them a more action-relevant retrieval signal than raw image or language features alone.' Validated empirically in Table 3 but not proven.
  • domain assumption Robot manipulation tasks exhibit recurring visual states and task phases that induce similar conditional transport paths toward structured action chunks.
    Invoked in §1: 'Robot manipulation often involves recurring visual states, task phases, and language goals, which can induce similar conditional transport paths.' This is the empirical premise that makes caching useful.
  • domain assumption Sparse ternary random projection preserves enough similarity structure for effective nearest-neighbor retrieval.
    Invoked in §3.2 via reference to [18]. The Johnson-Lindenstrauss property is standard, but its adequacy for this specific retrieval task is assumed, not proven.
  • domain assumption Actions from successful episodes are valid warm-start candidates for future similar contexts.
    Invoked in §3.3: 'If the episode completes successfully, these pending pairs are committed to the main cache for future reuse.' Assumes success implies action quality.

pith-pipeline@v1.1.0-glm · 16906 in / 3368 out tokens · 411546 ms · 2026-07-08T08:03:46.966225+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action (VLA) models have emerged as a promising approach for generalizable robotic manipulations. In particular, flow matching-based VLA models have shown remarkable success due to their capability to generate precise and smooth action sequences and capture multimodal distributions. However, the iterative denoising process in the action head acts as a major computational bottleneck, posing a critical challenge for real-time deployment. To address this challenge, we propose ActionCache, a plug-and-play external cache that opportunistically reuses past intermediate actions to warm-start generations from the vicinity of target actions, thereby drastically reducing the inference latency. Specifically, ActionCache stores the intermediate actions with compact multimodal keys, which enables retrieval from similar past contexts across different episodes or even different tasks. Experimental results in simulation and real-world environments demonstrate that ActionCache maintains high task success rates in a low-latency regime, achieving inference acceleration of up to $11.75\times$ and $34.43\times$ for representative flow-based VLA models, $\pi_{0.5}$ and GR00T-N1.6, respectively.

Figures

Figures reproduced from arXiv: 2607.06370 by Daichi Fujiki, Hikari Otsuka, Kosuke Matsushima, Masato Motomura, Ryuji Oi, Tatsuya Kaneko, Yuki Ichikawa.

Figure 1
Figure 1. Figure 1: Inference latency break￾down for representative VLA models on NVIDIA RTX 4090 GPU. Action head accounts for 37- 66% of the end-to-end latency. As shown in Equation (1), generating each action chunk in a flow-based VLA requires multi-step flow integration, where the action head repeatedly evaluates the velocity field Vθ. Thus, the number of flow matching steps directly determines the number of action head f… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ActionCache. ActionCache is an external cache that stores intermediate noisy actions from past generations. When a cache hit occurs, ActionCache initializes from retrieved past similar actions, thereby reducing the number of denoising steps. 3.1 Framework Overview [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Success rate and latency trade-off for ActionCache and baselines. (a): ActionCache achieves higher success rates throughout all NFEs compared to the baselines. (b): As the Nhit decreases, the latency of the action head decreases proportionally with a negligible latency overhead. "Seen" object configurations. We measure wall-clock latencies for each model/method on a single NVIDIA RTX 5090 GPU with 32 GB of… view at source ↗
Figure 4
Figure 4. Figure 4: Cache hit rate across episode timesteps. Cache hit rate reaches over 80% at episode start stage. To investigate the potential for cross-task action reuse, we conduct a zero-shot cross-task experiment on VLABench with π0.5. In this experiment, we first populate the cache with the select_fruit task and then evaluate the model’s performance on other target tasks. For a more fine-grained analysis of action reu… view at source ↗
Figure 5
Figure 5. Figure 5: (a): Average Top-1 similarity vs success rate for various numbers of denoising steps. As the Top-1 similarity increases, the success rates improve. (b): Average Top1 similarity vs hit thresholds for various cache size. As the hit threshold increases, average Top-1 similarity also increases. (c): Cache hit rate vs various hit thresholds for various cache size. A Larger cache yields a higher cache hit rate. … view at source ↗
Figure 6
Figure 6. Figure 6: Hit rate comparison for cache replacement policies un￾der a restricted cache size of 100. LFU policy demonstrates the high￾est hit rate across hit thresholds. Comparison of Cache Replacement Policies under Re￾stricted Capacity. As established in Section 4.4, the quality of retrieved actions is important to achieve a high success rate in ActionCache. With a large cache size, we can retrieve suffi￾ciently hi… view at source ↗
Figure 7
Figure 7. Figure 7: Real-world trajectories and corresponding cache hit rates. The cache hit rate is high on average during the approaching and placing phases, while it drops during the grasping phase. The video can be found in the supplementary material. specific and difficult grasps, LFU prioritizes robust and highly reusable actions that are shared across tasks or episodes. Although LFU captures temporal locality less effe… view at source ↗
Figure 8
Figure 8. Figure 8: Real-world experimental environment. We use 6-DoF SO-101 robotic arm with an overhead and wrist-mounted camera, a box and a white cube. B Results for LIBERO benchmark LIBERO consists of four task suites (Spatial, Object, Goal and Long) which contain 10 tasks each. We evaluate the base π0.5 model on LIBERO benchmark and report success rates for each task suites at NFE = 1 and 10. We evaluate on a total 200 … 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

35 extracted references · 35 canonical work pages · 7 internal anchors

  1. [1]

    Approximate caching for efficiently serving text-to-image diffusion models

    Shubham Agarwal, Subrata Mitra, Sarthak Chakraborty, Srikrishna Karanam, Koyel Mukherjee, and Shiv Kumar Saini. Approximate caching for efficiently serving text-to-image diffusion models. InUSENIX Symposium on Networked Systems Design and Implementation. 2024

    work page 2024

  2. [2]

    L. A. Belady. A study of replacement algorithms for a virtual-storage computer.IBM Systems Journal. 1966

    work page 1966

  3. [3]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, Edith LLontop, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed, You Liang Tan, Guanzhi...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Robert Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, brian ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren,...

    work page 2025

  5. [5]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.π0: A Visio...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    JAX: composable transformations of Python+NumPy programs

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Yash Katariya, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman- Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs. 2018

    work page 2018

  7. [7]

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

    work page 2023

  8. [8]

    LeRobot: An open-source library for end-to-end robot learning

    Remi Cadene, Simon Alibert, Francesco Capuano, Michel Aractingi, Adil Zouitine, Pepijn Kooijmans, Jade Choghari, Martino Russi, Caroline Pascal, Steven Palma, Dana Aubakirova, Mustafa Shukor, Jess Moss, Alexander Soare, Quentin Lhoest, Quentin Gallouédec, and Thomas 10 Wolf. LeRobot: An open-source library for end-to-end robot learning. InInternational Co...

    work page 2026

  9. [9]

    Falcon: Fast visuomotor policies via partial denoising

    Haojun Chen, Minghao Liu, Chengdong Ma, Xiaojian Ma, Zailin Ma, Huimin Wu, Yuanpei Chen, Yifan Zhong, Mingzhi Wang, Qing Li, and Yaodong Yang. Falcon: Fast visuomotor policies via partial denoising. InInternational Conference on Machine Learning. 2025

    work page 2025

  10. [10]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems. 2023

    work page 2023

  11. [11]

    Real-time iteration scheme for diffusion policy

    Yufei Duan, Hang Yin, and Danica Kragic. Real-time iteration scheme for diffusion policy. In International Conference on Intelligent Robots and Systems. 2025

    work page 2025

  12. [12]

    VLA-0: Building state-of-the-art vlas with zero modification.arXiv preprint arXiv:2510.13054

    Ankit Goyal, Hugo Hadfield, Xuning Yang, Valts Blukis, and Fabio Ramos. VLA-0: Building state-of-the-art vlas with zero modification.arXiv preprint arXiv:2510.13054. 2025

    work page arXiv 2025

  13. [13]

    Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems. 2020

    work page 2020

  14. [14]

    Fast Policy Synthesis with Variable Noise Diffusion Models

    Sigmund H Høeg, Yilun Du, and Olav Egeland. Fast Policy Synthesis with Variable Noise Diffusion Models. InInternational Conference on Robotics and Automation. 2025

    work page 2025

  15. [15]

    Vision-language- action models for robotics: A review towards real-world applications.IEEE Access

    Kento Kawaharazuka, Jihoon Oh, Jun Yamada, Ingmar Posner, and Yuke Zhu. Vision-language- action models for robotics: A review towards real-world applications.IEEE Access. 2025

    work page 2025

  16. [16]

    OpenVLA: An open-source vision-language-action model

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, Thomas Kollar, Ben- jamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. OpenVLA: An open-source vision-language-action model. InConference on Robot Learning. 2024

    work page 2024

  17. [17]

    STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction

    Jinhao Li, Yuxuan Cong, Yingqiao Wang, Hao Xia, Shan Huang, Yijia Zhang, Ningyi Xu, and Guohao Dai. STEP: Warm-started visuomotor policies with spatiotemporal consistency prediction.arXiv preprint arXiv:2602.08245. 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  18. [18]

    Hastie, and Kenneth W

    Ping Li, Trevor J. Hastie, and Kenneth W. Church. Very sparse random projections. InSIGKDD Conference on Knowledge Discovery and Data Mining. 2006

    work page 2006

  19. [19]

    CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation

    Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, Xiaofan Wang, Bei Liu, Jianlong Fu, Jianmin Bao, Dong Chen, Yuanchun Shi, Jiaolong Yang, and Baining Guo. CogACT: A foundational vision-language- action model for synergizing cognition and action in robotic manipulation.arXiv preprint arXiv:2...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Vision-language foundation models as effective robot imitators

    Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, Hang Li, and Tao Kong. Vision-language foundation models as effective robot imitators. InInternational Conference on Learning Representations. 2024

    work page 2024

  21. [21]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InInternational Conference on Learning Representations. 2023

    work page 2023

  22. [22]

    LIBERO: Benchmarking knowledge transfer for lifelong robot learning

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, qiang liu, Yuke Zhu, and Peter Stone. LIBERO: Benchmarking knowledge transfer for lifelong robot learning. InNeural Information Processing Systems Datasets and Benchmarks Track. 2023

    work page 2023

  23. [23]

    DeepCache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. InComputer Vision and Pattern Recognition. 2024

    work page 2024

  24. [24]

    A survey on vision– language–action models for embodied AI.IEEE Transactions on Neural Networks and Learning Systems

    Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King. A survey on vision– language–action models for embodied AI.IEEE Transactions on Neural Networks and Learning Systems. 2026

    work page 2026

  25. [25]

    PyTorch: an imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: an imperative style, high-performa...

    work page 2019

  26. [26]

    SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, Simon Alibert, Matthieu Cord, Thomas Wolf, and Remi Cadene. SmolVLA: A vision-language-action model for affordable and efficient robotics.arXiv preprint arXiv:2506.01844. 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control

    Junjie Wen, Yichen Zhu, Jinming Li, Zhibin Tang, Chaomin Shen, and Feifei Feng. DexVLA: Vision-language model with plug-in diffusion expert for general robot control.arXiv preprint arXiv:2502.05855. 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    TinyVLA: Toward fast, data-efficient vision-language-action models for robotic manipulation.IEEE Robotics and Automation Letters

    Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Zhibin Tang, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, and Jian Tang. TinyVLA: Toward fast, data-efficient vision-language-action models for robotic manipulation.IEEE Robotics and Automation Letters. 2025

    work page 2025

  29. [29]

    MoDM: Efficient serving for image generation via mixture-of-diffusion models

    Yuchen Xia, Divyam Sharma, Yichao Yuan, Souvik Kundu, and Nishil Talati. MoDM: Efficient serving for image generation via mixture-of-diffusion models. InInternational Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1. 2025

    work page 2025

  30. [30]

    Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys. 2023

    work page 2023

  31. [31]

    EfficientVLA: Training-free acceleration and compression for vision- language-action models

    Yantai Yang, Yuhao Wang, Zichen Wen, Luo Zhongwei, Chang Zou, Zhipeng Zhang, Chuan Wen, and Linfeng Zhang. EfficientVLA: Training-free acceleration and compression for vision- language-action models. InAdvances in Neural Information Processing Systems. 2025

    work page 2025

  32. [32]

    Vision-language models for vision tasks: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence. 2024

    work page 2024

  33. [33]

    VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks

    Shiduo Zhang, Zhe Xu, Peiju Liu, Xiaopeng Yu, Yuan Li, Qinghui Gao, Zhaoye Fei, Zhangyue Yin, Zuxuan Wu, Yu-Gang Jiang, and Xipeng Qiu. VLABench: A large-scale benchmark for language-conditioned robotics manipulation with long-horizon reasoning tasks.arXiv preprint arXiv:2412.18194. 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    DreamVLA: A vision-language-action model dreamed with comprehensive world knowledge

    Wenyao Zhang, Hongsi Liu, Zekun Qi, Yunnan Wang, XinQiang Yu, Jiazhao Zhang, Runpei Dong, Jiawei He, He Wang, Zhizheng Zhang, Li Yi, Wenjun Zeng, and Xin Jin. DreamVLA: A vision-language-action model dreamed with comprehensive world knowledge. InAdvances in Neural Information Processing Systems. 2025

    work page 2025

  35. [35]

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

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R Sanketi, Grecia Salazar, Michael S Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, ...

    work page 2023