Robotics · arXiv:2609.34911 v2 · submitted 28 Sep 2026 · 14 references
Don't Throw Away the Tail: Action Upcycling for Policy Acceleration
The paper says
“Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency.”
more of the abstract
A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose Action Upcycling, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2-1.7x with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 30 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A genuinely useful, training-free acceleration with a new signal source and a sound adaptive-vs-fixed ablation, but the 'no loss in success' headline is unverifiable at the reported single-seed and 10-episode resolution. This line is wrong
4 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 14 references say.
The paper says, at §4.1.1, Tables 1–5
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
All simulation success rates are reported as single numbers with no variance across seeds, and the claimed gains are sub-point or 1–2 point deltas (e.g., SmolVLA 82.6→82.8, π0.5 96.9→97.9, RoboTwin 2.0 π0.5 59.7→60.5). At this resolution, the headline 'no loss in success rate' is indistinguishable from 'no measurable change within sampling noise.'
more of objection 1
The paper should report mean and standard deviation over at least three to five seeds and apply a significance test (or paired comparison) before claiming that success is preserved.
The paper says, at §4.2, Tables 6–7
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The real-robot evidence consists of 10-episode counts per task. GR00T N1.6 goes 34/40→35/40 overall (a one-episode difference) and the per-task changes are 8/10→9/10; π0.5's drawer task changes of 7/10→9/10 and 8/10→10/10 are two-episode differences. None of these deltas is resolvable at the 10-episode sample size, so the real-world 'matches or improves on every task' claim (Section 4.2.1) is not supported at the stated resolution.
more of objection 2
Reporting the per-episode call counts and timing deltas is useful, but the success-rate claim needs more episodes or a statistical treatment.
The paper says, at §3.3, Eq. (2), Algorithm 1; §B.2, Figure 6
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The reported 1.2–1.7× call reduction is largely a construction artifact: τ is selected from the pool so that the mean execution length satisfies h̄(τ) ≥ r·h (Algorithm 1, line 2), so the call reduction follows by definition from the per-policy target r.
more of objection 3
The genuinely empirical content is the success-preservation claim, and that is the weakly supported part. Figure 6 further shows that success is maintained only for r in a narrow band (roughly 1.2–1.7) and drops beyond it, with the real-world deployments using r=1.25 and r=1.1. The paper does acknowledge this in Section B.2, but the abstract and conclusion state 'no loss in success rate' without the r-conditioning, which overstates what is demonstrated.
The paper says, at §3.2, Figure 1
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The safety margin of the threshold is never directly measured. Figure 1's correlation between c_k and tail/replanned RMSE is an aggregate over all rollout steps and is reported on LIBERO only, whereas the deployment-relevant quantity is the conditional distribution of RMSE among exactly the tail actions admitted at c_k ≤ τ.
more of objection 4
A direct check would be: among steps where hexec > h, what is the RMSE of the executed tail actions versus the replanned ones, and how does its distribution compare with the baseline prefix? Without this, the claim that low velocity fluctuation is a task-robust proxy for tail-action validity rests on unstated assumptions about monotonicity and threshold calibration on new tasks.
4 minor objections
The paper says, at §3.2, Eq. (3)
Pith says: minor objection 1 of 4
The velocity convention (v_k = a_k for relative displacements, v_k = a_k − a_{k-1} for absolute positions) is stated in prose but should be written out as a definition with a subscript convention, since the choice affects the scale of c_k and hence the pool-based threshold τ.
The paper says, at Table 1
Pith says: minor objection 2 of 4
Latency is reported at integer-ms precision and seconds at two decimals; given that several success-rate deltas are sub-point, the tables should also report the number of evaluation episodes per cell and whether the LIBERO figure is the standard 50-episode protocol.
show 2 more
The paper says, at §4.1.2, Table 5
Pith says: minor objection 3 of 4
The fixed-length control is a good idea, but the fixed-length success rates (97.05, 81.55, 95.40, 96.45) are themselves single-seed numbers, so the 0.9–1.3-point adaptive-vs-fixed gap is also within the resolution concern raised above; this should be rechecked with variance reported.
The paper says, at §A.4
Pith says: minor objection 4 of 4
The real-world baseline execution horizons (h=15 for π0.5, h=9 for GR00T N1.6) are stated to have been chosen because they 'gave the most stable execution in our preliminary trials'; the criteria for that choice should be specified so the comparison is reproducible.
What would settle it
Re-run the Table 1 and Tables 6/7 evaluations with at least five independent seeds (and at least 50 episodes per real-robot task) to place 95% confidence intervals on every success-rate delta, and in the same runs record, for every chunk admitted at the deployed τ, the RMSE between the upcycled tail actions and the replanned actions the baseline would have executed; report both the CIs and the 95th-percentile admitted RMSE normalized by the median inter-step displacement. If the upcycled and baseline CIs overlap for most cells, or if the admitted RMSE at the 95th percentile is large relative to typical step size, the no-loss claim is noise-limited and the threshold's safety margin is thin.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 30 Sep 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Short version: this is a real, useful contribution with one overstated headline. The velocity-fluctuation signal, computed from a single action chunk with no model internals and no extra samples, is genuinely new relative to the attention- and sampling-based adaptive-horizon methods it cites. The method is clean, training-free, and composes with few-step sampling (2.8x) and FlashVLA (6.7x). I believe the core mechanism.
What earns credit: the adaptive-versus-fixed-length ablation (Table 5) is the right control, and it lands. Fixed-length execution degrades success in three of four policies; adaptive execution beats fixed-length in all four. That consistent direction supports the claim that adaptivity, not just longer execution, is doing the work. Figure 1's correlation (0.89-0.98) motivates the signal, the pool-construction analysis (2% of chunks suffices, online pool works) is a thoughtful detail, and the code is released with public checkpoints used in simulation.
The soft spots, in proportion. First, the statistics: every simulation cell is a single seed, and the real-robot cells are ten-episode counts. GR00T's 34/40 to 35/40 is one episode; SmolVLA's 82.6 to 82.8 on LIBERO is well inside sampling noise. 'No loss in success rate' cannot be distinguished from 'no measurable change' at this resolution. Second, the call reduction is largely a restatement of the tuned ratio r: Eq. 2 and Algorithm 1 set the threshold so the mean execution length reaches r*h. The paper is transparent about this, since Figure 6 sweeps r and shows success holds only in the 1.2-1.7 band, but the honest reading is that the empirical content is 'success retained in a hand-tuned band,' not 'no loss.' Third, the threshold's safety margin is never directly measured. Figure 1's RMSE-versus-fluctuation correlation is an aggregate over rollouts, reported on LIBERO only; deployment safety depends on the conditional distribution of RMSE among exactly the tail actions admitted at ck <= tau, which is uncharacterized.
None of this sinks the mechanism. The pattern is consistent across policies and the ablation is supportive, so this is a conditional verdict, not a rejection. But the headline claim is soft at the stated resolution, and the paper would be materially stronger with seeds and variance reported and an explicit statement of how success varies across r within the working band. Worth citing, worth discussing, and worth sending to review with a request for those experiments.
The strongest objection
The headline claim decomposes into two parts. The call-reduction factor (1.2–1.7×) is a restatement of the chosen ratio r: by Eq. 2, r = h-bar/h, and the threshold τ is selected from the pool so that h-bar(τ) ≥ rh (Algorithm 1, Section 3.3), so the call reduction holds by construction. The genuinely empirical content is 'no loss in success rate,' and that claim is verified at a resolution the experiments cannot support. Every simulation cell in Table 1 is a single seed (e.g., SmolVLA 82.6→82.8, π0.5 96.9→97.9), and the real-robot cells in Tables 6 and 7 are 10-episode counts: GR00T N1.6 overall 34/40→35/40 is a one-episode difference, drawer baseball→basket 7/10→9/10 is two episodes. None of these deltas exceeds the sampling noise the protocol can resolve, and the paper itself acknowledges in Section B.2 / Figure 6 that success drops outside a narrow hand-tuned band of r (1.2–1.7), with real-world deployments using r=1.25 and r=1.1. Moreover, the safety margin of the velocity-fluctuation threshold is never directly measured: Figure 1's correlation between c_k and tail/replanned RMSE is an aggregate over rollouts and is reported on LIBERO only, whereas deployment safety depends on the conditional distribution of RMSE among exactly the tail actions admitted at c_k ≤ τ. The one comparison that isolates the mechanism, the adaptive-vs-fixed ablation (Table 5), shows ~1-point single-seed differences. The result is that 'no loss in success' cannot be distinguished from 'no measurable change,' and generalization to untested policies and tasks rests on an uncharacterized safety margin.
The paper's strongest genuine asset is the adaptive-vs-fixed ablation (Table 5): at matched mean execution length, adaptive thresholding beats fixed-length execution by 0.9–1.3 points on all four policies, which is real evidence that the c_k signal carries information beyond merely lengthening the horizon. Released code, public checkpoints, no extra samples, and no model-internal access are also real strengths, and the real-robot deployment is a useful falsification risk that the authors chose to take. My concern is not about the mechanism's plausibility but about the resolution at which the central empirical claim is verified. Because the call reduction is a definitional consequence of the chosen r (Eq. 2) and the only measured quantity is success retention, the load-bearing evidence is the success rates, which are single-seed simulation deltas and 10-episode real counts. The paper's own Figure 6 acknowledges the tuned-band sensitivity, so I do not raise that as a new finding; instead I route the concern through correctness risk: the differences reported cannot resolve a loss or gain at the claimed level, and the c_k proxy's safety at the deployed threshold is never quantified, only its aggregate correlation on LIBERO. This reinforces rather than overturns the reader's CONDITIONAL verdict: the conditions for ACCEPT (multi-seed error bars and a characterized sensitivity to r) are exactly what is missing. I therefore recommend the verdict stay CONDITIONAL (UNCHANGED), and I partially agree with the reader's weakest_assumption: the reader emphasizes proxy reliability and the r band, while I sharpen the proxy point to the unmeasured conditional margin at τ and the noise-limited verification.
The referee's report
If the success-preserving claim holds, Action Upcycling is a genuinely useful and original addition to the policy-acceleration toolbox: it is training-free, requires no model internals and no extra sampling, applies to any chunked policy including a World Action Model, and composes with orthogonal acceleration methods. The manuscript's strengths are real: the mechanism is simple and internally coherent, the adaptive component is isolated by a sensible fixed-length control (Table 5), all simulation experiments use public checkpoints with released code, and the fixed-versus-adaptive ablation is the right kind of control. The main risk is that the paper's headline empirical claim ('no loss in success rate') is verified at a resolution the experiments cannot support: single-seed simulation cells and 10-episode real-robot counts. The call-reduction factor is, as the authors' own Eq. 2 and Algorithm 1 make clear, largely a restatement of the hand-chosen ratio r. The paper is therefore valuable as a method proposal but currently over-claims its empirical support.
Pith's reading of the paper (a machine's summary, not the authors')
Chunked robot policies predict a block of future actions but execute only a short prefix before re-observing and replanning; the rest of the block is discarded. This paper claims those discarded actions are worth keeping: they stay close to what the policy would plan next, so long as the predicted velocity within the chunk stays smooth. Action Upcycling measures accumulated velocity fluctuation from the single already-sampled chunk and extends execution up to the point where it rises, needing no model internals, no extra samples, and no training. Because the signal is free, the method cuts policy calls by 1.2–1.7x with no loss in success across three vision-language-action models and a world action model, and it composes with other accelerations — reaching 2.8x combined with few-step sampling and 6.7x with streaming decoding. A reader should care because it opens a new, nearly free axis of policy acceleration that multiplies with existing per-call speedups and held up on physical-robot drawer and pick-and-place tasks.
What carries the argument. The central object is the accumulated velocity fluctuation signal $c_k = \sum_{j=h+1}^{k} \|v_j - v_{j-1}\|_2$, computed only from the already-sampled action chunk, with velocity defined as the action itself for relative displacements and as the action difference for absolute joint positions. It is non-decreasing, stays small under steady motion, and rises as soon as velocity fluctuates. Action Upcycling executes tail actions while $c_k \le \tau$, and sets the threshold $\tau$ from a pool of such signals so that the mean execution length reaches $r\cdot h$ for a chosen upcycling ratio $r$ — this is what converts the signal into an adaptive horizon without accessing model internals, drawing extra samples, or training.
Core claim. The paper's central claim is that the tail of an action chunk — the actions a chunked policy predicts but does not execute — can be safely reused whenever the predicted motion stays smooth, because discarded actions track their replanned versions closely. Concretely, on the four tested policies the Pearson correlation between discarded and replanned actions is 0.89–0.98, and the RMSE between them grows with the accumulated change between consecutive predicted velocities. Action Upcycling turns this into a control rule: reuse tail actions while the accumulated velocity fluctuation $c_k$ stays under a threshold $\tau$, with $\tau$ set so that the mean execution length reaches a chosen upcycling ratio $r$. Across LIBERO, LIBERO-Plus, and RoboTwin 2.0, and in real-world pick-and-place and drawer tasks on a physical arm, this cuts policy calls by 1.2–1.7x while matching or slightly improving success (for example, $\pi_{0.5}$ on LIBERO goes from 96.9% to 97.9%), and the call reduction stacks with few-step sampling and streaming action decoding to reach 2.8x and 6.7x total speedups respectively.
The premise it rests on. The central bet is that low accumulated velocity fluctuation marks tail actions that will agree with replanning, a correlation established empirically on the four tested policies and tasks and only within a tuned upcycling-ratio band of about 1.2–1.7, beyond which the paper's own Figure 6 shows success falling.
If the paper is right (Pith's inferences, not the authors')
Any chunked policy can apply the method by computing the fluctuation signal from a single sampled chunk, making the acceleration model-agnostic and essentially free at inference time.
Because it reduces the number of policy calls rather than the cost per call, the gain multiplies with per-call accelerations: 2.8x combined with few-step sampling and 6.7x combined with streaming action decoding on $\pi_{0.5}$.
Reusing smooth tails is not merely lossless: success improves slightly on most evaluated settings, suggesting that longer open-loop execution can be mildly beneficial when the motion is steady.
The fixed-length ablation shows the benefit is not simply executing more actions — adaptive selection beats fixed-length execution at the same mean length by 0.9–1.3 success points across all four policies.
The headline "1.2–1.7x" is essentially the chosen upcycling ratio $r$: mean execution length is $r\cdot h$ by construction, so the real contribution of the signal is letting the policy operate at a higher $r$ without the success drop that fixed-length execution incurs — the honest reading is "advance up to the sweet spot the signal makes safe," not "the signal discovers a faster policy."
If smoothness genuinely tracks replanning agreement across embodiments, the same threshold logic should transfer to any chunked action-space policy, including whole-body or multi-agent policies, provided the velocity definition is re-derived for the action parameterization; that transfer is a testable extension the paper does not run.
Calibrating $\tau$ per policy from pool rollouts leaves a deployment hazard the paper only partially probes: an environment whose perturbation statistics differ from the calibration distribution could admit harmful tail actions at the same $r$, and the LIBERO-Plus results cover the seven perturbations actually tested rather than arbitrary distribution shift.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
How Pith checks papers · The full audit of this paper · Review history (2 versions)
The paper
Abstract
Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose Action Upcycling, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2-1.7x with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-29
The authors' main claim, as the classifier picked it out 0.97
“Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2--1.7$\times$ with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM).”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.64 · human participants 0.04 · controlled comparison 0.50 · randomized assignment 0.02 · states quantitative result 0.98 · compares with prior results 0.86 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. | background 1.00 | 0.07 | 0.03 | asserted 1.00 |
| Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. | background 0.99 | 0.10 | 0.03 | asserted 1.00 |
| A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. | background 1.00 | 0.08 | 0.02 | asserted 1.00 |
| Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. | gap 0.61 | 0.05 | 0.02 | asserted 1.00 |
| We propose *Action Upcycling*, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. | contribution 0.93 | 0.86 | 0.21 | asserted 1.00 |
| We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. | result 0.98 | 0.95 | 0.11 | asserted 0.99 |
| Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. | method 0.62 | 0.78 | 0.14 | asserted 0.98 |
| Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2--1.7$\times$ with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). | result 1.00 | 0.96 | 0.73 | asserted 1.00 |
| It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration. | interpretation 0.49 | 0.68 | 0.19 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-29: 1 request, 6,961 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (8)
Figures are the authors', reproduced from the paper.
References (14)
- [1] Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokin- sky, Shihao Cao, Thomas Charbonnier, et al.π0.7: a steerable generalist robotic foundation model with emergent capabilities.arXiv preprint arXiv:2604.15483, on PithPith's reviewarXiv
- [4] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xi- anliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and bench- mark with strong domain randomization for robust bimanual robotic manipulation.arXiv preprin... on PithPith's reviewarXiv
- [5] Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, et al. Libero-plus: In-depth robustness analysis of vision-language-action models.arXiv preprint arXiv:2510.13626, on PithPith's reviewarXiv
- [6] 10 Preprint. Filippo Lazzati, Kyle Stachowicz, William Chen, Alberto Maria Metelli, Andrew Wagenmaker, and Sergey Levine. Why does action chunking improve behavioral cloning performance in robotic control?arXiv preprint arXiv:2608.02547, on PithPith's reviewarXiv
- [7] Zekai Li, Jiaming Tang, and Zhijian Liu. Flashvla: Streaming action decoding for fast and asyn- chronous vla inference.arXiv preprint arXiv:2608.27384, on PithPith's reviewarXiv
- [8] •π 0.5 (Black et al., 2025a) is a VLA with a flow-matching action expert that attends to vision-language tokens at every layer through shared attention. We useN=10 denoising steps. For RoboTwin 2.0, we use the public checkpoint released by Motus (Bi et al., 2026). •SmolVLA(Shu... on Pith
- [10] URLhttps://openreview.net/forum?id=E2VsqgKNlr. Runze Xu, Xiaolong Shan, Shuang Dai, Yu Wang, and Jincheng Yu. Knowing when to stop: Adaptive action chunking via internal cross-attention dynamics in vlas.arXiv preprint arXiv:2609.00908, on PithPith's reviewarXiv
- [11] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666, on PithPith's reviewarXiv
- [12] Michael Zeng, Abhinav Agarwal, Ajay Bati, Brian Lee, Siddharth Ancha, and Russ Tedrake. Revis- iting open-loop execution in robotics: Toward reactive, higher-performing policies.arXiv preprint arXiv:2608.15938, on PithPith's reviewarXiv
- [13] doi: 10.15607/RSS.2023.XIX.016. 12 Preprint. A FURTHERIMPLEMENTATIONDETAILS A.1 POLICIES We apply Action Upcycling to the following policies using their released checkpoints without any fine-tuning for simulation experiments. The execution and prediction horizons of each polic... on Pithdoi:10.15607/rss.2023.xix.016
- [2023] Yuejiang Liu, Jubayer Hamid, Annie Xie, Yoonho Lee, Max Du, and Chelsea Finn. Bidirectional decoding: Improving action chunking via guided test-time sampling. InInternational Conference on Learning Representations, volume 2025, pp. 4594–4627, 2025a. Ziyan Liu, Yeqiu Chen, Hong... on PithPith's reviewarXiv
- [2024] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics.arXiv preprint arXiv:2506.01844, on PithPith's reviewarXiv
- [2025] 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 L... on PithPith's reviewarXiv
- [2026] Johan Bjorck, Fernando Casta ˜neda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, on PithPith's reviewarXiv
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment