Pith.

Robotics · arXiv:2611.03815 v1 · submitted 2 Nov 2026 · 40 references

AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data

Yueh-Hua Wu, Kei Ota, Hanlong Li, Hayata Yuasa, Jason Wang, Kengo Ikeuchi, Kensuke Chiyoma, Makoto Sugiura

show 12 more authors

Masafumi Takahashi, Masatoshi Tateno, Naoaki Kanazawa, Nicola Scianca, Petr Khrapchenkov, Satoshi Otsubo, Takehiko Ohkawa, Takuya Okubo, Tatsuya Matsushima, Yuki Noguchi, Yuma Shin, Zixuan Zhong

The paper says

“AIRoA-0 (A-0) is a bimanual manipulation system learned from 16,151 hours of handheld demonstrations, with no robot data collected at any stage.”

more of the abstract

The demonstrations are recorded with YUBI, our open-source handheld gripper, which lets people record manipulation data anywhere with no robot present. From a single instruction, A-0 sorts more than a dozen small parts into their numbered cells and changes the mop pads and brush of a robot vacuum; among other tasks, it unzips a pencil case and folds a towel. It performs these tasks using only a camera on each wrist, with no head or scene camera, at a site where none of its training data was recorded. Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations. We collect no teleoperation data and use no reinforcement learning or on-robot corrections, and the robot appears only at test time. The system learns which regions of each wrist view matter for the task, predicts the stage the task has reached, and decodes actions as a continuous stream. A-0 is embodiment-agnostic: it runs on robots that mount the YUBI gripper, such as a bimanual Franka, TX-G2, OpenArm and UR. We are releasing the dataset, to our knowledge the largest open dataset of handheld-gripper demonstrations with task and subtask instructions, together with the pre-trained weights and the post-training recipe. Videos of A-0 on every task are on the project page, https://a-0.airoa.io. Anyone can build the gripper, record a task of their own and post-train the model on it in under four hours on a single node of eight H200 GPUs.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 4 Nov 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. Impressive system for learning from handheld data, but the headline success rate is undermined by underspecified partial credit scoring. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 40 references say.

The paper says, at Section 2.1; Table 1

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The evaluation protocol for partial credit is critically underspecified. The paper states that each success rate is 'an average over 15 trials, with partial credit for a task that is partly completed' but does not define (a) what milestones determine a binary success or partial credit for each task, (b) how partial credit is quantified (e.g., proportion of stages, a …

more of objection 1

… rubric, subjective rating), (c) whether the same partial-credit rubric was applied consistently to all baseline models, or (d) whether the evaluators were blind to model identity. Since the headline claim of 78% vs. 27% rests on these numbers, this lack of detail makes the central quantitative result unverifiable. The authors must provide a clear, objective scoring rubric for each task—ideally the one used by the evaluators—and state whether scoring was blinded.

The paper says, at Section 6.2; Figure 11

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The non-uniform streaming decoding requires a dedicated post-training procedure where the noise level per slot is drawn from a Beta-distributed interval (Eq. 6). The paper reports that this adds no parameters and that 'a streaming policy is obtained from the same pre-trained model with the same post-training recipe.'

more of objection 2

However, the post-training recipe (Table 4) does not mention any special training schedule for the streaming decoder—e.g., how the two-slot buffer and the Beta(1.5,1) sampling are integrated into the 20,000-step training loop. It should be clarified whether the post-training data is augmented with simulated buffer states or if the streaming behavior emerges purely from the noise-level conditioning. Without this, the claim that the same recipe works for both standard and streaming decoding is not fully supported.

The paper says, at Section 4.6; Figures 9, 10

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The scaling analysis pre-trains on fractions of the full corpus (100%, 50%, etc.) but keeps all other factors (batch size, schedule) identical. The power-law fit (L = 0.063 D^{-0.035}) shows a very weak dependence on data size—a doubling of data yields only a ~2.4% loss reduction.

more of objection 3

While the authors interpret this as a positive scaling signal, it should be acknowledged that the benefit is small and may saturate quickly. A brief discussion of whether this weak scaling is expected for handheld data or whether it constrains the practical value of further data collection would strengthen the analysis.

5 minor objections

The paper says, at Section 3.2

Pith says: minor objection 1 of 5

The corpus is described as covering 'well over a hundred tasks', but a precise count or a distribution of task types would help readers assess diversity.

The paper says, at Section 4.2, Eq. (1)

Pith says: minor objection 2 of 5

The anchored action representation is defined for each arm separately. It would be helpful to clarify whether the 20 numbers per action include both arms concatenated or are per-arm predictions.

show 3 more

The paper says, at Section 5.1, Table 3

Pith says: minor objection 3 of 5

The stage prediction head is trained with cross-entropy. What is the average number of stages per task across the ten evaluation tasks? This context would help gauge the difficulty of the classification.

The paper says, at Section 2.3

Pith says: minor objection 4 of 5

The cross-embodiment results on TX-G2, OpenArm, and UR are presented for a subset of tasks (2–4 tasks per robot). The number of trials per task for these robots is not stated, nor is any success rate given. Including even a small number of quantitative results would strengthen the embodiment-agnostic claim.

The paper says, at Figure 8

Pith says: minor objection 5 of 5

The visual budget gate's attention maps are shown for 25 subtasks. The figure would benefit from a brief caption explaining the contour levels and confirming whether these are aggregated over multiple timesteps or shown for a single frame.

What would settle it

Publish the exact scoring rubric for each task, including the stage-level binary pass/fail conditions, the partial credit formula (e.g., fraction of stages completed without exceeding a timeout), and whether the same rule applies to all models. Then, have two independent annotators blind to model identity re-score the recorded evaluation trials (or a random subset of 5 trials per task) using the published rubric. If the inter-annotator agreement (Cohen's κ) is below 0.8, or if re-scored success rates differ from the reported ones by more than 5 percentage points, the evaluation protocol is unreliable and the reported numbers require revision.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 4 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 0 s · 16,994 tokens read, 5,777 written
  3. Editor wrote the verdict above 22 s · 17,870 tokens read, 1,333 written
  4. Referee major revision: 3 major objections, 5 minor 37 s · 18,603 tokens read, 1,541 written

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

This paper delivers on a promising direction: a complete pipeline from handheld demonstrations to real robot control, with no robot data at any stage. The scale is real (16,151 hours), the cross-embodiment results on four robot types are a genuine demonstration of generality, and the technical components—visual evidence budget, stage prediction, non-uniform streaming decoding—are sensibly integrated. The baseline comparison is thorough: each baseline gets a hyperparameter search, while A-0 uses a single recipe. The scaling law from pre-training data is a nice touch. If these results hold up, this is a significant step toward robot learning that scales with human effort, not robot fleet size.

Now the soft spots. The biggest is the evaluation protocol for the central claim of 78% success. The paper says 'partial credit for a task that is partly completed' but gives no rubric for what constitutes a complete vs. partial trial, how partial credit is quantified, whether the same standard is applied to baselines, or whether the scoring was blind. That is a load-bearing gap for a paper that leads with this number. The stress-test concern lands: without this information, the 78% vs. 27% comparison is not independently verifiable. This is a flaw that needs to be fixed before I'd rely on the headline result.

Other soft spots: the evaluation is mostly on one platform (Franka) with 15 trials per task—moderate sample size. The other robots get only qualitative demos. The post-training recipe is claimed as 'universal' but the hyperparameters were presumably tuned on these tasks; claiming universality without cross-scene or cross-task validation (e.g., on a new task recorded at a third site) is premature. Code and data are promised but not yet released, which matters for a system billed as open.

On the plus side, the paper is honest about limitations: it notes the data is all human, the robot never sees failure recoveries, the wrist-only view is partial. The architecture choices are clearly motivated and ablation experiments (scale, streaming schedule) are helpful.

For whom? Anyone working on scaling robot learning via human data collection, especially from handheld grippers. The system-level design is a useful reference. The paper deserves a serious referee, but the evaluation protocol needs to be disclosed before the quantitative claims can be accepted at face value.

The strongest objection

The paper's central quantitative result is a 78% average success rate across 10 tasks (Section 2.1, Table 1). The evaluation methodology is described as 'each an average over 15 trials, with partial credit for a task that is partly completed.' No further detail is provided: (1) For each task, what objective milestones define a binary success? (2) How is partial credit quantified (e.g., proportion of stages completed, or a subjective rating)? (3) Is the same partial credit rubric applied to all baseline models? (4) Who performed the scoring and were they blind to the model identity? These missing details create ambiguity about whether the reported advantage over baselines (78% vs 27% for π0.5) reflects genuine superiority or differences in scoring. Since the paper positions this comparison as its headline evidence, the evaluation protocol is a load-bearing component that is currently insufficiently specified for reproducible verification.

The paper's main quantitative claim is the 78% success rate across ten tasks, compared to 27% for the best baseline. While the project page shows videos, the numerical comparison is central to establishing superiority. The lack of a precise evaluation protocol for partial credit is a concrete, load-bearing weakness that can be addressed by the authors through documentation and independent re-scoring. This concern reinforces the need for clarification that the reader already flagged under CONDITIONAL acceptance; it does not change the verdict direction. The reader's weakest assumption focused on transfer without calibration and stage generalization, but this complementary concern about evaluation rigor is equally important for the central claim.

The referee's report

If the reported results are reproducible and the evaluation is sound, this work represents a significant advance: it shows that large-scale handheld data alone can yield a system capable of long-horizon, precise bimanual manipulation that transfers to multiple robot embodiments without any robot-specific data or calibration. The cross-embodiment demonstration, the one-recipe post-training, and the open release of data and models would make this a highly impactful contribution to robot learning.

Pith's reading of the paper (a machine's summary, not the authors')

AIRoA-0 (A-0) is a bimanual robot manipulation system that learns solely from 16,151 hours of handheld demonstrations recorded with the YUBI gripper, requiring no robot data, teleoperation, or reinforcement learning. The paper claims that from a single instruction and through only two wrist cameras, A-0 can perform precise, multi-minute tasks such as sorting parts, folding towels, and servicing a robot vacuum at a site where none of its training data was recorded. Across ten evaluation tasks, A-0 averages 78% success, far above the best baseline (π0.5 at 27%). The system is embodiment-agnostic, running on multiple robot arms (Franka, TX-G2, OpenArm, UR) with the same weights, and it recovers from disturbances like moved bins by relying on what it sees.

What carries the argument. The central objects are (1) anchored actions defined as relative SE(3) poses from the gripper at the time of observation, which makes handheld and robot trajectories share the same coordinate system; (2) the visual evidence budget gate, which learns per-token keep probabilities from the control objective alone, highlighting task-relevant regions; (3) a stage head that predicts the current subtask stage from wrist views, conditioning the action expert via language; and (4) non-uniform streaming decoding that holds a buffer of two action chunks at noise levels (0.1, 1), refining the chunk about to be executed with only a small correction on the newest observation to keep motion continuous.

Core claim. The paper shows that a complete bimanual manipulation policy can be learned from handheld data alone, without any robot-specific data, reinforcement learning, or on-robot corrections. The central claim is that the combination of an anchored action representation (relative SE(3) poses), a visual evidence budget gate that learns which regions of the wrist camera views matter, stage prediction to track long-horizon tasks, and non-uniform streaming decoding for continuous motion, all built on a pre-trained VLA backbone (π0.5), yields a system that generalizes to an unseen test site and to different robot embodiments. The policy's success is demonstrated across ten diverse bimanual tasks, with a 78% average success rate vs. 27% for the best baseline. The paper further claims that the pre-trained model scales with data: more handheld demonstrations produce lower action loss both in pre-training and after post-training.

The premise it rests on. The assumption that the anchored action representation (relative SE(3) poses) and the visual budget gate, both trained solely on handheld demonstrations, transfer directly to robots at an unseen test site without any robot-specific calibration, domain adaptation, or post-training on in-domain data, and that the stage prediction learned from subtask-labeled data generalizes to task variations not covered in the training set.

If the paper is right (Pith's inferences, not the authors')

If correct, A-0 demonstrates that robot manipulation policies can be trained exclusively from human handheld demonstrations, which scale with the number of people recording, not the number of robots.

The system's embodiment agnosticism means that a single policy can be deployed on different robot arms without retraining, as long as they mount the same gripper and wrist cameras.

The non-uniform streaming decoder enables continuous real-time action generation with only two denoising evaluations per query (60 ms on an H200), making it practical for closed-loop control.

The stage prediction and visual budget gate together allow the policy to track long-horizon tasks and ignore irrelevant visual variation across deployment sites, contributing to robust zero-shot site transfer.

The paper leaves implicit that the same approach could be applied to tasks beyond the ten evaluated, but the 30% success on the earbuds task suggests the method may struggle with exceptionally fine or complex manipulations; future work could probe whether data diversity or a different action representation helps.

The power-law scaling from pre-training data (L ∝ D⁻⁰·⁰³⁵) is shallow but reliable; a reader might infer that even larger handheld datasets (e.g., 100,000 hours) would yield diminishing returns, and that the marginal gain per hour may be small.

The system's reliance on the YUBI gripper's exact wrist-camera placement means that any change to the camera mount or optics would require re-collection of training data, a limitation not discussed.

The visual budget gate is trained solely from the action loss; the paper could have tested whether explicit attention maps from a segmentation model would match or exceed its performance, but no such comparison is made.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

AIRoA-0 (A-0) is a bimanual manipulation system learned from 16,151 hours of handheld demonstrations, with no robot data collected at any stage. The demonstrations are recorded with YUBI, our open-source handheld gripper, which lets people record manipulation data anywhere with no robot present. From a single instruction, A-0 sorts more than a dozen small parts into their numbered cells and changes the mop pads and brush of a robot vacuum; among other tasks, it unzips a pencil case and folds a towel. It performs these tasks using only a camera on each wrist, with no head or scene camera, at a site where none of its training data was recorded. Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations. We collect no teleoperation data and use no reinforcement learning or on-robot corrections, and the robot appears only at test time. The system learns which regions of each wrist view matter for the task, predicts the stage the task has reached, and decodes actions as a continuous stream. A-0 is embodiment-agnostic: it runs on robots that mount the YUBI gripper, such as a bimanual Franka, TX-G2, OpenArm and UR. We are releasing the dataset, to our knowledge the largest open dataset of handheld-gripper demonstrations with task and subtask instructions, together with the pre-trained weights and the post-training recipe. Videos of A-0 on every task are on the project page, https://a-0.airoa.io. Anyone can build the gripper, record a task of their own and post-train the model on it in under four hours on a single node of eight H200 GPUs.

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 · 2026-11-04

The authors' main claim, as the classifier picked it out 0.88

“Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.12 · states limitation 0.25 · human participants 0.64 · controlled comparison 0.92 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.97

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.

SentenceRoleThe authors' own workStates the main conclusionStrength
AIRoA-0 (A-0) is a bimanual manipulation system learned from 16,151 hours of handheld demonstrations, with no robot data collected at any stage. contribution 0.72 0.92 0.12 asserted 1.00
The demonstrations are recorded with YUBI, our open-source handheld gripper, which lets people record manipulation data anywhere with no robot present. method 0.95 0.87 0.04 asserted 1.00
From a single instruction, A-0 sorts more than a dozen small parts into their numbered cells and changes the mop pads and brush of a robot vacuum; among other tasks, it unzips a pencil case and folds a towel. result 0.92 0.90 0.17 asserted 1.00
It performs these tasks using only a camera on each wrist, with no head or scene camera, at a site where none of its training data was recorded. method 0.59 0.89 0.08 asserted 1.00
Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations. result 1.00 0.95 0.31 asserted 1.00
We collect no teleoperation data and use no reinforcement learning or on-robot corrections, and the robot appears only at test time. method 0.93 0.94 0.07 asserted 1.00
The system learns which regions of each wrist view matter for the task, predicts the stage the task has reached, and decodes actions as a continuous stream. method 1.00 0.86 0.07 asserted 1.00
A-0 is embodiment-agnostic: it runs on robots that mount the YUBI gripper, such as a bimanual Franka, TX-G2, OpenArm and UR. contribution 0.74 0.71 0.10 asserted 1.00
We are releasing the dataset, to our knowledge the largest open dataset of handheld-gripper demonstrations with task and subtask instructions, together with the pre-trained weights and the post-training recipe. contribution 1.00 0.95 0.07 asserted 0.83
Videos of A-0 on every task are on the project page, https://a-0.airoa.io. contribution 0.93 0.70 0.03 asserted 1.00
Anyone can build the gripper, record a task of their own and post-train the model on it in under four hours on a single node of eight H200 GPUs. contribution 0.80 0.51 0.05 asserted 0.83

What kind of paper the classifier reports this is: method or tool 0.98 · evidence from observation or experiment 1.00

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, 2026-11-04: 1 request, 8,414 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 (12)
Figure 1 of the paper
Figure 1. A-0 in overview. Handheld demonstrations, seen here through their wrist cameras, are the only data we collect. The four parts of the system turn them into actions for robots that mount the same gripper and cameras; the photographs show towel folding on the Franka, the pencil case on TX-G2, colour sorting on OpenArm and the orange pyramid on UR. The chart gives each model’s success rate, averaged over the ten tasks of Section 2.1. one of its steps. Data that anyone can record is the work of many operators, here across well over a hundred tasks, and a single model has to learn from…
Figure 2 of the paper
Figure 2. The ten evaluation tasks, each in a frame from an evaluation run of A-0
Figure 3 of the paper
Figure 3. Two minutes of one autonomous run of the robot-vacuum task, with the time on the robot since (a). (a) The gripper takes hold of the first used mop pad, and (b) the second comes off. (c) A new pad is carried from the box and (d) pressed onto its mount; (e) both new pads are on, and (f) the brush is pulled out. 4
Figure 4 of the paper
Figure 4. The bins moved during autonomous colour sorting, with the time on the robot since (a). (a) An arm carries a green part to the green bin; the bins are, from left to right, yellow, green and red. (b) A person pushes the bins to the left, and the red bin is now under the gripper. (c) The arm moves over to the green bin and (d) drops the part into it. 2.3. Other robots Nothing in A-0 refers to the arm that carries the grippers: state and action are defined at the grippers themselves (Section 4.2), and the model sees only through the wrist cameras. A robot that takes the YUBI gripper…
Figure 5 of the paper
Figure 5. A-0 on TX-G2, OpenArm and UR, each task in a frame from a run on that robot. 3. Data The rest of this report describes how A-0 is built, in four parts designed with one another in view (
Figure 6 of the paper
Figure 6. A-0 as one system. Handheld demonstrations pass through data processing into pre-training; post-training adapts the pre-trained model to the tasks it will perform; the inference harness runs it on the robot. 3.1. YUBI YUBI [2] is a handheld gripper built for collecting manipulation data at scale. It is driven by the fingers, so that closing the hand closes the jaws; it carries a camera at the wrist; and it is tracked in six degrees of freedom. A demonstrator holds one gripper in each hand and performs the task as they naturally would, and the recording holds the two wrist views,…
Figure 7 of the paper
Figure 7. The A-0 model. Pre-training adds the visual budget gate and subtask prediction to the 𝜋0.5 backbone. Post-training adds the stage head, whose prediction conditions the action expert as text, and the proprioceptive memory. 7
Figure 8 of the paper
Figure 8. Where the visual budget gate puts its weight, on 25 subtasks. Each tile shows the left and right wrist views the model receives. The contours trace the smoothed keep probability at three levels, and the view fades where the gate gives it less weight. where 𝜏 is a temperature and 𝛾 a floor that keeps a trace of every token. A budget loss holds the mean keep probability over the 𝑃 tokens of a view to a target 𝜌: ℒbudget = ( 1 𝑃 𝑃 ∑ 𝑝=1 𝑚𝑝 − 𝜌)2 . (3) The budget constrains the average gate weight, while the action loss teaches the gate which regions to emphasise when the model…
Figure 9 of the paper
Figure 9. Action loss on held-out demonstrations for models pre-trained on a share of the data. Left, during pre-training, with the task instruction as prompt (the subtask instruction gives the same picture): every run has the same model, batch size and schedule, and the 100% run is the pre-training of A-0. Right, during post-training, averaged over 16 task sets: each run starts from its checkpoint at 100,000 pre-training steps, and post-training data and settings are the same for all. 1k 2k 4k 8k 16k 0.045 0.046 0.047 0.048 0.049 0.050 100% 50% 25% 12.5% 6.25% Pre-training data (hours)…
Figure 10 of the paper
Figure 10. Post-training action loss on held-out demonstrations against the amount of pre-training data, on logarithmic axes. Each point is the 16-set average in the right panel of
Figure 11 of the paper
Figure 11. Streaming decoding with two slots. Left: the buffer before and after one query. Right: the noise level of one chunk on its way through the buffer, with a uniform schedule and with the non-uniform schedule used in A-0. With evenly spaced levels, as in FlashVLA, the size of every update is tied to the number of slots. With few slots the update just before execution is large, and on the robot this shows as seams between consecutive chunks; with many, each chunk is settled on observations taken long before it runs, and the policy is slow to respond when the scene changes. We propose…
Figure 12 of the paper
Figure 12. gives the post-training loss in the right panel of

Figures are the authors', reproduced from the paper.

References (40)
  1. [1] C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake and S. Song. Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots. Robotics: Science and Systems , 2024 on Pith
  2. [2] T. Ohkawa, J. Arima, Y. Noguchi, M. Tateno, M. Sugiura, T. Okubo, K. Ikeuchi, Y. Shin, H. Nishizawa, N. Kanazawa et al. YUBI: Yielding Universal Bidigital Interface for Bimanual Dexterous Manipulation at Scale. arXiv:2606.10244, 2026 on PithPith's reviewarXiv
  3. [3] Y.-H. Wu et al. See Less, Specify More: Visual Evidence Budgets for Generalizable VLAs. Conference on Robot Learning, 2026. 15 AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data on Pith
  4. [4] Physical Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess et al. 𝜋0.5: a Vision-Language- Action Model with Open-World Generalization. arXiv:2504.16054, 2025 on PithPith's reviewarXiv
  5. [5] J. Guo, P. Jin, J. Li, P. Li, Y. Li, F. Liu et al. Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories. arXiv:2607.15330, 2026 on PithPith's reviewarXiv
  6. [6] NVIDIA. Isaac GR00T N1.7. Hugging Face model nvidia/GR00T -N1.7-3B and GitHub repository NVIDIA/Isaac- GR00T, accessed October 2026 on Pith
  7. [7] J. Bjorck, F. Castañeda, N. Cherniadev , X. Da, R. Ding, L. Fan et al. GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734, 2025 on PithPith's reviewarXiv
  8. [8] H. Fang, J. Duan, D. Clay , S. Wang, S. Liu, W. Huang et al. MolmoAct2: Action Reasoning Models for Real-world Deployment. arXiv:2605.02881, 2026 on PithPith's reviewarXiv
  9. [9] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn et al. 𝜋0: A Vision-Language-Action Flow Model for General Robot Control. arXiv:2410.24164, 2024 on PithPith's reviewarXiv
  10. [10] Z. Li, J. Tang and Z. Liu. FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference. arXiv:2608.27384, 2026 on PithPith's reviewarXiv
  11. [11] N. M. M. Shafiullah, A. Rai, H. Etukuru et al. On Bringing Robots Home. arXiv:2311.16098, 2023 on PithPith's reviewarXiv
  12. [12] H. Etukuru, N. Naka, Z. Hu et al. Robot Utility Models: General Policies for Zero-Shot Deployment in New Environments. arXiv:2409.05865, 2024 on PithPith's reviewarXiv
  13. [13] F. Lin, Y. Hu, P. Sheng et al. Data Scaling Laws in Imitation Learning for Robotic Manipulation. International Conference on Learning Representations , 2025 on Pith
  14. [14] S. Liu, B. Li, K. Ma, L. Wu, H. Tan, X. Ouyang, H. Su and J. Zhu. RDT2: Exploring the Scaling Limit of UMI Data Towards Zero-Shot Cross-Embodiment Generalization. arXiv:2602.03310, 2026 on PitharXiv
  15. [15] Generalist Team. GEN-1: Scaling Embodied Foundation Models to Mastery . Generalist AI Blog, 2026 on Pith
  16. [16] Reward AI Team. OM-1: Frontier Robot Intelligence, Learned Firsthand from Humans. Reward AI Blog, 2026 on Pith
  17. [17] Y. Wei, J. Ma, J. Wang, W. Zhou, Y. Zuo, K. Rui et al. HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone. arXiv:2607.25895, 2026 on PithPith's reviewarXiv
  18. [18] GenRobot. 10Kh RealOmni-Open DataSet. Hugging Face dataset genrobot2025/10Kh-RealOmin-OpenData, ac- cessed October 2026 on Pith
  19. [19] A. Brohan, N. Brown, J. Carbajal et al. RT -2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. Conference on Robot Learning , 2023 on Pith
  20. [20] Octo Model Team, D. Ghosh, H. Walke, K. Pertsch et al. Octo: An Open-Source Generalist Robot Policy . Robotics: Science and Systems , 2024 on Pith
show 20 more
  1. [21] M. J. Kim, K. Pertsch, S. Karamcheti et al. OpenVLA: An Open-Source Vision-Language-Action Model. Conference on Robot Learning , 2024 on Pith
  2. [22] Open X-Embodiment Collaboration. Open X-Embodiment: Robotic Learning Datasets and RT -X Models. arXiv:2310.08864, 2023 on PithPith's reviewarXiv
  3. [23] A. Khazatsky , K. Pertsch, S. Nair et al. DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset. Robotics: Science and Systems , 2024 on Pith
  4. [24] Gemini Robotics Team. Gemini Robotics: Bringing AI into the Physical World. arXiv:2503.20020, 2025 on PithPith's reviewarXiv
  5. [25] Generalist Team. GEN-0: Embodied Foundation Models That Scale with Physical Interaction. Generalist AI Blog, 2025 on Pith
  6. [26] Physical Intelligence, A. Amin, R. Aniceto, A. Balakrishna et al. 𝜋∗0.6: a VLA That Learns From Experience. arXiv:2511.14759, 2025 on PithPith's reviewarXiv
  7. [27] S. Belkhale, T. Ding, T. Xiao et al. RT -H: Action Hierarchies Using Language. Robotics: Science and Systems , 2024 on Pith
  8. [28] L. X. Shi, B. Ichter , M. Equi et al. Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language- Action Models. International Conference on Machine Learning , 2025 on Pith
  9. [29] I. Larchenko, G. Zarin and A. Karnatak. Task Adaptation of Vision-Language-Action Model: 1st Place Solution for the 2025 BEHAVIOR Challenge. arXiv:2512.06951, 2025. 16 AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data on PitharXiv
  10. [30] M. Torne, A. Tang, Y. Liu and C. Finn. Learning Long-Context Diffusion Policies via Past-Token Prediction. Con- ference on Robot Learning , 2025 on Pith
  11. [31] A. J. Hancock, A. Z. Ren and A. Majumdar . Run-time Observation Interventions Make Vision-Language-Action Models More Visually Robust. arXiv:2410.01971, 2024 on PithPith's reviewarXiv
  12. [32] T. Z. Zhao, V. Kumar , S. Levine and C. Finn. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Robotics: Science and Systems , 2023 on Pith
  13. [33] C. Chi, Z. Xu, S. Feng et al. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Robotics: Science and Systems, 2023 on Pith
  14. [34] K. Black, M. Y. Galliker and S. Levine. Real-Time Execution of Action Chunking Flow Policies. Advances in Neural Information Processing Systems, 2025 on Pith
  15. [35] B. Chen, D. Martí Monsó, Y. Du et al. Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. Advances in Neural Information Processing Systems , 2024 on Pith
  16. [36] S. H. Høeg, Y. Du and O. Egeland. Streaming Diffusion Policy: Fast Policy Synthesis with Variable Noise Diffusion Models. arXiv:2406.04806, 2024. 17 AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data A. Post-training Loss by Task Set Figure 12 giv... on PithPith's reviewarXiv
  17. [37] 10 Articulated objects 0.05 0.06 0.07 Box packing 0.06 0.08 Carrier packing 0.06 0.08 on Pith
  18. [38] 10 Clothing handling 0.045 0.050 0.055 Electrical connection 0.06 0.08 Holder organization 0.04 0.05 0.06 Industrial kitting 0.035 0.040 0.045 Letter arrangement 0.08 on Pith
  19. [39] 12 Mechanical assembly 0.04 0.06 0.08 Robot vacuum service 0.04 0.05 0.06 Selection sorting 0.04 0.05 Stacking balance 0 1 2 3 4 5 0.04 0.05 Tableware serving 0 1 2 3 4 5 0.08 on Pith
  20. [40] 12 Textile folding 0 1 2 3 4 5 0.04 0.05 0.06 Throwing 0 1 2 3 4 5 0.05 0.06 0.07 Transfer Post-training update (thousands) Held-out action loss Pre-training data 100% 50% 25% 12.5% 6.25% Figure 12 | The loss in the right panel of Figure 9 for each of the 16 task sets. 18 on Pith

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.

No comments yet.

Sign in to comment