Pith.

Robotics · arXiv:2609.35652 v1 · submitted 28 Sep 2026 · 119 references

MM-ABC: Towards Generalist Mobile Manipulation via Seeing, Coordinating and Imagining

Qiwei Liang, Guangyu Chen, Shaolong Zhu, Zikuan Xiao, Jinxuan Lu, Yifan Xie, Renjing Xu, Wenbo Ding

show 1 more author

Tianxing Chen

The paper says

“Mobile manipulation extends robot interaction beyond a fixed kinematic workspace by making the reachable region itself controllable. This flexibility introduces two central challenges: spatially grounded perception under continuous ego-motion and coordinated control of heterogeneous arm and base actions.”

more of the abstract

Existing approaches strengthen geometry through explicit 3D representations or predictive world models, and often decouple mobility and manipulation into separate action streams. We argue that effective mobile manipulation requires not only decoupling, but also representations that support efficient cross-stream collaboration. We present MM-ABC, a foundation model built around Seeing, Coordinating, and Imagining Arm-Base Collaboration. MM-ABC combines sparse multi-level VLM features for spatial perception; a training-only future branch that uses world imagination and geometric intent as extra supervision, strengthening perception and manipulation-intent prediction and improving the overall learning signal; and MM-APT, which coordinates separate manipulation and mobility streams through masked joint attention and clean-action x-prediction. In controlled ablations, replacing clean-action prediction with velocity prediction lowers success on RoboCasa365 composite-seen tasks from 32.8% to 29.2%, and removing future supervision or multilevel conditioning causes larger drops. We pretrain MM-ABC on 5,000+ hours of heterogeneous robot data spanning 400K+ episodes, 12 datasets, and 17 embodiments. Experiments cover EBench, RoboCasa365, ManiSkill-HAB, LIBERO, LIBERO-Plus, and real-world mobile manipulation. MM-ABC achieves 44.71% success on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 29 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. Solid systems paper with honest, well-controlled ablations, but its signature clean-action mechanism is load-bearing on a synthetic premise that the real-data ablation supports only weakly. 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 119 references say.

The paper says, at Section 3, Eq. (3); Section 6.2, Table 6

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The quantitative claims of Section 3 (23x lower endpoint variance, 1.64x excess risk, 14% lower task-allocation error) are all computed inside the generator of Eq. (3), which prescribes a rank-13 action subspace of which 8 latent dimensions are deterministic functions of the conditioning c0 and feeds the arm–body allocation m to both heads.

more of objection 1

Under this construction the x-head can emit the conditional mean from context alone near t=0 while the v-head must cancel injected noise, so the generator largely dictates the reported gap. The transfer to MM-ABC requires real 64x80 canonical action chunks to be comparably low-dimensional with a large conditioning-determined share, and requires that learning to infer the allocation (which the authors acknowledge is given in the synthetic task but inferred in MM-ABC, end of Section 3) does not change the ranking. The paper provides no spectral evidence for real chunks: Figure 1(a) shows only the synthetic spectrum, and the only real-data bridge is the from-scratch RoboCasa365 ablation (32.8% vs 29.2%, a +3.6-point single-run comparison with no seeds or variance, whose per-task effects are heterogeneous, e.g., v-pred wins WashLettuce by 15 points while x-pred wins PreSoakPan by 15.5). Since the authors have 5000+ hours of real canonicalized action data at hand, they should either report the effective rank or participation ratio of real action chunks to test the low-rank premise, or run the component ablation over several seeds and report variance; without one of these, the clean-action contribution is asserted on a premise the paper never verifies on its own data.

The paper says, at Section 6.2, Table 6

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The ablation as reported supports the clean-action claim only in the full-model configuration, and the cross-variant comparison points the other way on most tasks. Comparing the two rows that each remove exactly one component from the full model, x-pred (clean action, future supervision removed) averages 24.9% while v-pred (velocity prediction, future supervision retained) averages 29.2%; v-pred is higher on 12 of the 16 tasks, for example SteamInMicrowave 27.0 vs 10.0, SetUpCuttingStation 28.0 vs 18.0, and PackIdentical 7.0 vs 1.5.

more of objection 2

The paper's narrative (abstract, Section 6.2, Section 7) states that replacing clean-action with velocity prediction lowers success, which is true only of the full-vs-v-pred comparison, and never discusses the apparent interaction between the prediction target and future supervision. At minimum, this indicates the 3.6-point effect is not an independent, additive advantage of clean-action prediction, and a reader comparing the two removal rows could read the opposite conclusion. The paper should state this cross-comparison explicitly, report whether the full-vs-v-pred 3.6-point gap persists across seeds, and temper the general claim in the abstract and conclusion accordingly.

The paper says, at Section 6.1, Section 6.3

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The headline comparative margins are single point estimates with no reported variance or evaluation counts: EBench 44.71% vs 41.41% (+3.30), RoboCasa365 61.2% vs 54.2% (+7.0), and the real-world 83% vs 71% over 20 trials per task. For the real-world results, success rates of 75–90% on 20 trials have wide binomial confidence intervals, so the 12-point margin over pi0.5 is not clearly significant under a conventional error bar, and the paper gives no per-task trial counts or CIs.

more of objection 3

For EBench and RoboCasa365, no statement of seeds or evaluation episodes is provided. Given that the smallest claimed margins are close to typical stochastic evaluation noise in this field, the authors should report the number of evaluation episodes, and where feasible a small number of seeds or a binomial CI, for the comparisons that carry the 'exceeds baselines' claims.

5 minor objections

The paper says, at Figure 2

Pith says: minor objection 1 of 5

The diagram labels the joint transformer blocks 'MM-JiT' while the text and contribution list call the module MM-APT; the diagram should be renamed for consistency.

The paper says, at Abstract and Section 1

Pith says: minor objection 2 of 5

Several rendering errors remove spaces between words, such as 'WepresentMM-ABC' and 'seea changing workspace'; the final version should be re-proofed.

show 3 more

The paper says, at Section 6.2

Pith says: minor objection 3 of 5

The variant definitions deserve a sentence making explicit that x-pred removes future supervision while v-pred retains it, so that the two removal rows are not immediately comparable; this directly relates to the interaction issue raised above.

The paper says, at Figure 1(a)

Pith says: minor objection 4 of 5

The caption states the clean spectrum terminates at rank 13, as prescribed by the generator; since this is synthetic data, the caption should say so explicitly to avoid a reader mistaking it for real action-chunk evidence.

The paper says, at Section 5.4

Pith says: minor objection 5 of 5

The data-engine section is thorough but would benefit from reporting the fraction of raw hours rejected by each tier, so the reader can gauge the cleaning burden; Table 6 also omits per-variant training-time cost, which would help interpret the 7.9-point future-supervision effect.

What would settle it

Reproduce Fig. 1(a) on real actions: center and stack canonicalized 64x80 chunks from the pretraining corpus and the RoboCasa365 composite-seen demonstrations (per embodiment/task), and report normalized singular spectra for clean chunks x, flow velocity v, and injected noise epsilon, with the participation ratio and the rank needed for 99% of variance (synthetic reference: 13/928, about 1.4%). If real clean spectra decay comparably sharply, the low-rank premise lands and the Section 3 mechanism transfers; if the clean spectrum approaches full rank, the mechanism does not transfer and the real case rests on the ablation alone. As a secondary companion check, rerun the v-pred vs x-pred pair of Table 6 with at least five seeds and report the gap distribution; a 3.6-point mean whose interval spans zero would leave the clean-action component unsupported on real tasks.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 29 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 2 min 54 s · 30,071 tokens read, 20,305 written
  3. Editor wrote the verdict above 25 s · 31,058 tokens read, 3,477 written
  4. Referee major revision: 3 major objections, 5 minor 1 min 11 s · 32,483 tokens read, 2,015 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

You should know two things before reading this. It is a serious systems contribution to mobile manipulation VLA, not a hype piece: the 5,000+-hour, 17-embodiment pretraining corpus, the staged data engine, and the evaluation across EBench, RoboCasa365, ManiSkill-HAB, LIBERO/Plus, and five real-world tasks are all real and handled with unusual care. The second thing: the headline mechanism claim is only weakly supported on real data, and the authors appear to know it.

What is actually new is the MM-APT two-stream action transformer with masked joint attention over arm and body streams, and the transfer of clean-action x-prediction into that coupled setting. The components are honestly credited: clean-action prediction already appears in DP3, RDT-1B, ManiCM, ABot-M0, and VGFM; removable future-supervision branches in PHR-VLA, WAM4D, and MECo-WAM; multi-depth conditioning builds on DeepStack. The contribution is the combination in a two-stream coordination architecture, plus the MM-30 dataset and the synthetic study.

The strongest part is the ablations. From-scratch, common 120k-step budget, batch 64, five variants. Future supervision is worth 7.9 points, multilevel conditioning 7.2–10.8, clean-action prediction 3.6. The synthetic study is also careful: paired seeds, known Bayes-optimal denoiser, ridge probes. LIBERO-Plus at 82.8% without perturbation training is a genuine result.

Now the soft spot, and here I agree with the reader and the stress-test. Section 3's generator prescribes a rank-13 action subspace where 8 of 13 dimensions are deterministic functions of the conditioning c0, and feeds the arm–body allocation m to both heads as input. With no raw-input skip in either decoder, the x-head can read the conditional mean from context near t=0 while the v-head must cancel injected noise it cannot resolve — that is where the 23x, 1.64x, and 14% numbers come from. Transfer to real 64x80 chunks requires them to be comparably low-rank and requires that learning to infer the allocation does not change the ranking. The paper concedes the allocation is inferred, not given, in MM-ABC, and offers no spectral evidence for real chunks. The real-data bridge is a single-run +3.6-point ablation whose per-task effects are heterogeneous: v-pred wins WashLettuce by 15 points and ties RinseSinkBasin. Directionally consistent, but the magnitude is uncalibrated. On the positive side, the headline gaps over baselines are also modest — 3.3 points on EBench, 0.5 on LIBERO — so the case rests heavily on the ablations.

None of this is fatal. The direction is right and the paper is honest about the concession. But the mechanism claim overreaches its evidence. An updated version with seeds or variance on headline numbers, and either spectral analysis of real chunks or an architecture variant without the no-skip condition, would close the gap. This deserves a serious referee — it is a within-subfield results paper for people building mobile-manipulation policies, and I would send it out with a request for exactly that evidence.

The strongest objection

MM-APT's signature design choice — clean-action x-prediction over velocity prediction — is justified by the Section 3 controlled study, whose quantitative claims (23x lower endpoint-noise variance at t=0.01, 1.64x excess-risk ratio, 14% lower task-allocation error) are all computed inside the generator of Eq. (3), which prescribes a known rank-13 action subspace and feeds the arm-body allocation m to both heads as an input. Because 8 of the 13 latent dimensions are deterministic functions of the conditioning c0, the x-head can emit the conditional mean from context alone near t=0, while the v-head must cancel injected noise it cannot resolve from its finite-width features — the mechanism behind the gap. Transfer to MM-ABC requires real canonical action chunks (64x80) to be comparably low-rank with a large conditioning-determined share, and requires that learning to infer the allocation does not change the ranking. The paper provides no spectral evidence for real chunks (Fig. 1(a) shows only the synthetic spectrum) and concedes the allocation is inferred, not given, in MM-ABC (end of Section 3). The only real-data bridge is the from-scratch RoboCasa365 ablation (29.2% vs 32.8%, +3.6 points, Table 6), a single-run comparison with no seeds or variance, whose per-task effects are heterogeneous (v-pred wins WashLettuce by 15.0 points, ties RinseSinkBasin, and the gap is concentrated in a few tasks, e.g., PreSoakPan -15.5). A +3.6-point effect is directionally consistent with the synthetic study but far smaller than its implied magnitude, so the strength of the mechanism on real action distributions is uncalibrated: the low-rank premise is load-bearing and untested.

The paper's central claim has two pillars: the headline benchmark results and the three component contributions from the from-scratch ablations. The clean-action pillar is the least secure, for the same reason the reader identified, and I sharpen it in two ways. First, the Section 3 comparison does not merely assume low-dimensional actions; it assumes a rank-13 linear subspace of which 8 dimensions are fully determined by the conditioning, so the x-head's near-noise-free endpoint estimates at t=0.01 reflect context-following rather than denoising skill. What transfers depends on the effective rank and conditioning-determined share of real action chunks, both unmeasured; Fig. 1(a) is synthetic-only. Second, the real-data bridge is a single from-scratch ablation: Table 6's v-pred row differs by 3.6 points on average, with a win for v-pred on WashLettuce and per-task heterogeneity that no seed or variance information can contextualize. This is not fatal: the paper honestly concedes the allocation-inference gap (end of Section 3), states the no-raw-input-skip design explicitly (Fig. 1(b) caption and Section 4.3), and the from-scratch ablation protocol is otherwise well controlled across the five variants (Section 6.2). Classifier sentences b30-b31 and b6 state the synthetic and ablation evidence, and both are reported accurately relative to the full text. The risk is also bounded because clean-action prediction has independent precedent in DP3, RDT-1B, ABot-M0, and JiT (Section 2.3), so failure of the transfer premise would weaken the paper's mechanistic justification rather than invalidate the design choice. I considered the missing error bars on EBench and the 20-trial real-world protocol as alternative candidates; these are real reporting gaps but standard for the field and would only shrink margins, not reverse directions, so they are less load-bearing than an untested premise at the center of one of the three named components. A spectral analysis of real action chunks plus a multi-seed replication of Table 6 would settle the point. Because the concern is addressable and does not overturn the empirically reported benchmark or ablation measurements, the CONDITIONAL verdict stands unchanged.

The referee's report

If the results hold, the paper is a substantial systems contribution to mobile manipulation: a genuinely large and heterogeneous pretraining corpus (5000+ hours, 17 embodiments), an architecture that achieves state-of-the-art or near-state-of-the-art numbers across five simulation benchmarks plus a real-world head-to-head against two strong baselines on identical demonstrations, and a clean methodological structure in which the component ablations are from scratch with a common budget rather than fitted to reproduce headline numbers. The synthetic study of Section 3 is, on its own terms, carefully executed: paired seeds, a shared loss and conditioning for both heads, and an analytical Bayes-optimal baseline. The paper is weaker as a scientific claim about the general principle of clean-action prediction: the Section 3 generator partially encodes the conclusion, the transfer premise to real action chunks is untested, and the real-data support is a single-run, heterogeneous ablation. The engineering result is likely sound even if the mechanism-level claim is overstated.

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

This paper asks how a mobile manipulation policy should represent the two heterogeneous action streams, arm and base, once they are decoupled. It argues that decoupling alone is not enough; the streams must exchange information through representations that respect their different control semantics, and that the output parameterization of flow-based action generation matters. Its central claim is that predicting the clean action endpoint (x-prediction) beats predicting flow velocity (v-prediction) in a two-stream transformer, because velocities force a finite-width network to carry and cancel injected noise while clean actions map directly onto a low-dimensional action manifold. The authors support this with a controlled synthetic study and a component ablation, then show that the full model, MM-ABC, reaches strong results on simulation benchmarks and real-world mobile tasks.

What carries the argument. The load-bearing mechanism is clean-action x-prediction embedded in a structured two-stream action transformer called MM-APT. Under rectified flow with noise $\epsilon$ and flow time $t$, the noisy chunk is $z_t = (1-t)\epsilon + tx$, and the network either predicts $x$ directly (clean action) or predicts velocity $v = x - \epsilon$; because neither decoder has a raw-input skip, the velocity head must propagate injected noise through its layers to reconstruct the endpoint, while the clean-action head maps straight to the low-dimensional action manifold. MM-APT keeps manipulation and mobility in separate streams with separate normalization and attention projections, couples them through masked joint attention with an asymmetric near-far visibility pattern, and conditions both on sparse multilevel VLM features injected as gated residuals from a handful of intermediate backbone layers. A training-only World Expert predicts future geometric features from a frozen teacher at horizons $n+32$ and $n+64$, masking those future tokens from the action streams and removing the branch at deployment, so geometric supervision shapes the shared perceptual representation without adding inference cost.

Core claim. The paper's central claim is that under rectified flow, clean-action x-prediction and velocity v-prediction are algebraically equivalent recipes for recovering the same sampled trajectory, but they impose different representational loads on a finite-width two-stream action expert, and x-prediction is the better target. The mechanism is geometric: clean action chunks occupy a low-dimensional subspace (rank 13 in the synthetic generator), while velocity and noise span the ambient space, so a velocity head must carry injected Gaussian noise through its layers to form its endpoint estimate, whereas a clean-action head can map directly toward the action manifold. In a controlled study with a known Bayes-optimal denoiser, x-prediction retains far less noise in its endpoint estimates at high flow times, denoises more accurately at high noise, and realizes the prescribed arm-body task allocation more faithfully under few-step sampling. The component ablation shows the effect transfers to a learned setting on RoboCasa365 composite-seen tasks, where replacing x-prediction with v-prediction lowers success from 32.8% to 29.2%, while removing future supervision lowers it by 7.9 points and weakening perceptual conditioning by 7.2 to 10.8 points.

The premise it rests on. The clean-action advantage rests on the premise, built into the synthetic generator, that robot action chunks are low-dimensional and that the dominant failure mode for a velocity head is having to carry injected noise through its no-skip decoder; the only real-data test of this transfer is the 3.6-point from-scratch RoboCasa365 ablation, which the authors themselves acknowledge is a smaller and noisier effect than the synthetic study suggests.

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

Clean-action prediction may be a generally useful design choice for any multi-stream action generator whose streams share finite width, not just arm-base mobile manipulation.

A training-only future-geometry branch offers a way to densify the learning signal for perception and manipulation-intent prediction without deploying a world model or demanding extra computation at inference.

The same architecture transfers from mobile to fixed-base settings, reaching 99.1% on LIBERO and 82.8% on LIBERO-Plus without perturbation training, suggesting shared perception and action representations work across body configurations.

Sparse multilevel conditioning beats both final-layer-only and per-layer interleaved conditioning in the from-scratch ablation, informing how future VLA action experts should ingest visual features from the backbone.

The synthetic-to-real transfer implicitly assumes real robot action chunks are similarly low-dimensional; a direct testable prediction is that adding a raw-input skip to the velocity decoder, or deliberately widening the action manifold, should shrink or reverse the 3.6-point gap.

The future-supervision branch shows the largest ablation drop (7.9 points) yet is the least analyzed; an extension question is whether the gain comes specifically from geometry-oriented targets or would appear with any dense auxiliary regression target.

Masking future tokens from action streams means the policy cannot condition execution on predicted geometry at test time; an inference-time variant that does consume predicted geometry might close the remaining composite-task gap on RoboCasa365.

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

Mobile manipulation extends robot interaction beyond a fixed kinematic workspace by making the reachable region itself controllable. This flexibility introduces two central challenges: spatially grounded perception under continuous ego-motion and coordinated control of heterogeneous arm and base actions. Existing approaches strengthen geometry through explicit 3D representations or predictive world models, and often decouple mobility and manipulation into separate action streams. We argue that effective mobile manipulation requires not only decoupling, but also representations that support efficient cross-stream collaboration. We present MM-ABC, a foundation model built around Seeing, Coordinating, and Imagining Arm-Base Collaboration. MM-ABC combines sparse multi-level VLM features for spatial perception; a training-only future branch that uses world imagination and geometric intent as extra supervision, strengthening perception and manipulation-intent prediction and improving the overall learning signal; and MM-APT, which coordinates separate manipulation and mobility streams through masked joint attention and clean-action x-prediction. In controlled ablations, replacing clean-action prediction with velocity prediction lowers success on RoboCasa365 composite-seen tasks from 32.8% to 29.2%, and removing future supervision or multilevel conditioning causes larger drops. We pretrain MM-ABC on 5,000+ hours of heterogeneous robot data spanning 400K+ episodes, 12 datasets, and 17 embodiments. Experiments cover EBench, RoboCasa365, ManiSkill-HAB, LIBERO, LIBERO-Plus, and real-world mobile manipulation. MM-ABC achieves 44.71% success on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks.

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 0.75

“MM-ABC achieves 44.71% success on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.19 · human participants 0.04 · controlled comparison 0.85 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.51 · 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.

SentenceRoleThe authors' own workStates the main conclusionStrength
Mobile manipulation extends robot interaction beyond a fixed kinematic workspace by making the reachable region itself controllable. background 1.00 0.05 0.03 asserted 1.00
This flexibility introduces two central challenges: spatially grounded perception under continuous ego-motion and coordinated control of heterogeneous arm and base actions. background 0.95 0.08 0.03 asserted 1.00
Existing approaches strengthen geometry through explicit 3D representations or predictive world models, and often decouple mobility and manipulation into separate action streams. background 0.96 0.04 0.03 asserted 0.96
We argue that effective mobile manipulation requires not only decoupling, but also representations that support efficient cross-stream collaboration. gap 0.93 0.13 0.35 asserted 0.84
We present MM-ABC, a foundation model built around Seeing, Coordinating, and Imagining Arm-Base Collaboration. contribution 0.96 0.87 0.15 asserted 0.99
MM-ABC combines sparse multi-level VLM features for spatial perception; a training-only future branch that uses world imagination and geometric intent as extra supervision, strengthening perception and manipulation-intent prediction and improving the overall learning signal; and MM-APT, which coordinates separate manipulation and mobility streams through masked joint attention and clean-action x-prediction. method 0.97 0.91 0.11 asserted 1.00
In controlled ablations, replacing clean-action prediction with velocity prediction lowers success on RoboCasa365 composite-seen tasks from 32.8% to 29.2%, and removing future supervision or multilevel conditioning causes larger drops. result 0.98 0.96 0.06 asserted 1.00
We pretrain MM-ABC on 5,000+ hours of heterogeneous robot data spanning 400K+ episodes, 12 datasets, and 17 embodiments. method 0.99 0.96 0.03 asserted 1.00
Experiments cover EBench, RoboCasa365, ManiSkill-HAB, LIBERO, LIBERO-Plus, and real-world mobile manipulation. method 0.69 0.89 0.03 asserted 0.95
MM-ABC achieves 44.71% success on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks. result 1.00 0.97 0.40 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · 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, 2026-09-29: 1 request, 7,813 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 (10)
Figure 1 of the paper
Figure 1. Clean-action versus velocity prediction under a common loss and a given arm–body allocation. (a) Clean chunks occupy a prescribed 13-dimensional subspace, whereas velocity and noise span the ambient space. (b) Variance of the endpoint estimate across noise draws with the clean chunk fixed. (c) Excess risk over the Bayes-optimal denoiser, as a 𝑣/𝑥 ratio. (d) Task-allocation error of sampled chunks against the number of Euler steps. (b) Noise in the endpoint estimate. We compare both heads through their endpoint estimates 𝑥ˆ, which live in the same space. Holding a clean chunk fixed,…
Figure 2 of the paper
Figure 2. MM-ABC architecture. Left: sparse VLM features condition MM-APT; a training-only future branch receives geometric supervision from frozen VGGT-Omega features. Middle: joint transformer blocks (MM-JiT in the diagram) combine stream-specific transformations with masked attention and read-only perceptual context. Right: action streams communicate within each segment, and far tokens additionally read near tokens. Future queries follow the same temporal ordering but remain isolated from actions. generates a 64-step normalized action chunk 𝑋 ∈ R 64×80, split into manipulation 𝑋𝑀 ∈ R 64×58…
Figure 3 of the paper
Figure 3. Overview of the MM-ABC multi-embodiment pretraining corpus. The corpus contains 5,166.2 cleaned hours from 12 datasets and 51 subsets, spanning 17 embodiments. Sector areas show training sampling probabilities aggregated by dataset; the surrounding panels illustrate environments and embodiments. (a) Pretraining data Real 80% Sim 20% (b) Embodiment profiles Hours 0 250 500 750 1000 949.4 18.4% ABC-130k 757.0 14.6% RoboCOIN 680.5 13.2% BEHAVIOR-1K 571.4 11.0% Hy-Embodied 0.5 465.8 9.0% AgiBotWorld EE 350.0 6.8% InternData-A1 340.5 6.6% RealSource World 339.6 6.6% Galaxea Open World…
Figure 4 of the paper
Figure 4. Pretraining data and embodiment composition. (a) Dataset hours and duration shares. (b) Arm configuration, mobility, and control frequency. Hardware platform. The platform comprises a HexFellow Trigger-A3 omnidirectional mobile base and two AgileX PiPER-X 6-DoF manipulators. The base repositions the robot in the horizontal plane, and the two arms perform object interactions, including bimanual handovers. Their active control channels map to the base and left- and right-arm slots of the 80D interface (
Figure 5 of the paper
Figure 5. MM-30: self-collected mobile manipulation dataset. Representative tasks, object and scene diversity, and episode counts per task. Role in pretraining. After the semantic audit, coordinate alignment, and quality filtering described in Section 5.4, all cleaned demonstrations from MM-30 are included in MM-ABC pretraining (
Figure 6 of the paper
Figure 6. MM-ABC data engine. Sources undergo semantic auditing, conversion with dataset-specific adapters, and four stages of quality filtering. Retained contiguous segments are encoded in the masked 80D interface, independently validated, and incorporated into the training mixture. Source audit and field semantics. For each source, we fix its revision and record its schema, camera streams, control frequency, and robot metadata. A dataset-specific frame contract specifies the physical meaning of each state/action field: units, action type, parent frame, TCP, rotation convention, gripper…
Figure 7 of the paper
Figure 7. Simulation benchmarks. Representative task scenes from LIBERO/LIBERO-Plus, RoboCasa365, EBench, and ManiSkill-HAB, covering fixed-base and mobile manipulation. 6 Experiments We organize our evaluation around three questions. First, how does MM-ABC compare with strong VLA and WAM baselines on mobile and fixed-base manipulation, including under distribution shifts? Second, how do sparse multilevel perceptual conditioning, clean-action prediction, and future geometric supervision contribute to performance? Third, how effectively does MM-ABC coordinate base motion and object interaction…
Figure 8 of the paper
Figure 8. EBench evaluation. Success rate (left) and stage-wise progress score (right) for MM-ABC and selected baselines [19, 20, 100–104]. Higher is better
Figure 9 of the paper
Figure 9. Real-world task execution. Representative stages of the five tasks, from top to bottom: Birthday Party Setup, Office Folder Arrangement, Kitchen Work, Industrial Parts Organization, and Chemistry Lab Operation. Within each task, frames progress from left to right. ■ Kitchen Work. The robot picks up a food container from the table and pours the food into a pot. It then puts the lid on the storage container, places the container into the cabinet, and closes the cabinet door. ■ Industrial Parts Organization. The robot places scattered industrial parts neatly into a parts bin, carries…
Figure 10 of the paper
Figure 10. Real-world task success. Success rate (%) over 20 trials per task for each method. ■ Chemistry Lab Operation. The robot picks up a test tube, moves to the laboratory bench behind it, pours the reagent into a beaker that already holds another reagent, and stirs until the two are thoroughly mixed. Evaluation protocol. We compare MM-ABC with 𝜋0.5 [20] and StarVLA-GR00T [118], fine-tuning every method on the same 200 teleoperated demonstrations per task. Each method is evaluated over 20 trials per task, and a trial counts as successful only when all steps of the task are completed. We…

Figures are the authors', reproduced from the paper.

References (119)
  1. [1] Jiahao Liu, Cui Wenbo, Zhongpu Xia, Yongliang Wang, Haoran Li, and Dongbin Zhao. Incom: Intent-driven perception and structured coordination for mobile manipulation.arXiv preprint arXiv:2602.23024, 2026 on PithPith's reviewarXiv
  2. [2] Xiangyu Zhu, Renjun Wu, Luzhou Ge, Jinyan Liu, and Xuesong Li. Geohat: Geometry-adaptive hybrid action transformer for mobile manipulation.arXiv preprint arXiv:2606.13394, 2026 on PithPith's reviewarXiv
  3. [3] Qiwei Liang, Boyang Cai, Minghao Lai, Sitong Zhuang, Tao Lin, Yan Qin, Yixuan Ye, Jiaming Liang, and Renjing Xu. Bootstrap dynamic-aware 3d visual representation for scalable robot learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p... on Pith
  4. [4] Ruisen Tu, Arth Shukla, Sohyun Yoo, Xuanlin Li, Junxi Li, Jianwen Xie, Hao Su, and Zhuowen Tu. Sg-vla: Learning spatially-grounded vision-language-action models for mobile manipulation.arXiv preprint arXiv:2603.22760, 2026 on PitharXiv
  5. [5] Ruijie Zheng, Jing Wang, Scott Reed, Johan Bjorck, Yu Fang, Fengyuan Hu, Joel Jang, Kaushil Kundalia, Zongyu Lin, Loic Magne, et al. Flare: Robot learning with implicit world modeling.arXiv preprint arXiv:2505.15659, 2025 on PithPith's reviewarXiv
  6. [6] Ronghan Chen, Yandan Yang, Zuojin Tang, Dongjie Huo, Tong Lin, Haoning Wu, Haoyun Liu, Yuzhi Chen, Lulu Zheng, Botai Yuan, et al. Abot-m0. 5: Unified mobility-and-manipulation world action model.arXiv preprint arXiv:2607.00678, 2026 on PithPith's reviewarXiv
  7. [7] Ying Li, Xiaobao Wei, Jiajun Cao, Hao Wang, Xiaowei Chi, Chengyu Bai, Qianpu Sun, Jiajun Li, Xiaojie Zhang, Peidong Jia, et al. Wam4d: Fast 4d world action model via spatial register tokens. arXiv preprint arXiv:2606.14048, 2026 on PithPith's reviewarXiv
  8. [8] SixiangChen,JiamingLiu,SiyuanQian,HanJiang,ZhuoyangLiu,ChenyangGu,XiaoqiLi,Chengkai Hou, Pengwei Wang, Zhongyuan Wang, et al. Ac-dit: Adaptive coordination diffusion transformer for mobile manipulation.Advances in Neural Information Processing Systems, 38:64008–64036, 2026 on Pith
  9. [9] Guangyu Chen, Qiwei Liang, Shaolong Zhu, Tianxing Chen, Zikuan Xiao, Yifan Xie, Lingfeng Zhang, Ping Luo, Renjing Xu, and Wenbo Ding. Mopa: Coordinated mobile manipulation via subsystem-specific perception alignment.arXiv preprint arXiv:2609.12081, 2026 on PithPith's reviewarXiv
  10. [10] Oussama Khatib. Mobile manipulation: The robotic assistant.Robotics and Autonomous Systems, 26 (2-3):175–183, 1999 on Pith
  11. [11] Jiaheng Hu, Peter Stone, and Roberto Martín-Martín. Causal policy gradient for whole-body mobile manipulation.arXiv preprint arXiv:2305.04866, 2023 on PithPith's reviewarXiv
  12. [12] KianaEhsani,TanmayGupta,RoseHendrix,JordiSalvador,LucaWeihs,Kuo-HaoZeng,KunalPratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, et al. Spoc: Imitating shortest paths in simulation enables effective navigation and manipulation in the real world. In2024 IEEE/CVF Conference on... on Pith
  13. [13] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation.arXiv preprint arXiv:2401.02117, 2024. 21 on PithPith's reviewarXiv
  14. [14] Yunfan Jiang, Ruohan Zhang, Josiah Wong, Chen Wang, Yanjie Ze, Hang Yin, Cem Gokmen, Shuran Song, Jiajun Wu, and Li Fei-Fei. Behavior robot suite: Streamlining real-world whole-body manipulation for everyday household activities.arXiv preprint arXiv:2503.05652, 2025 on PithPith's reviewarXiv
  15. [15] Qiwei Liang, Boyang Cai, Rongyi He, Hui Li, Tao Teng, Haihan Duan, Changxin Huang, and Runhao Zeng. Whole-body coordination for dynamic object grasping with legged manipulators. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 18434–18442, 2026 on Pith
  16. [16] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023 on PithPith's reviewarXiv
  17. [17] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language- action model.arXiv preprint arXiv:2406.09246, 2024 on PithPith's reviewarXiv
  18. [18] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, andShuranSong. Diffusionpolicy: Visuomotorpolicylearningviaactiondiffusion.TheInternational Journal of Robotics Research, 44(10-11):1684–1704, 2025 on Pith
  19. [19] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.𝜋0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024 on PithPith's reviewarXiv
  20. [20] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, AdnanEsmail, MichaelEqui,ChelseaFinn, NiccoloFusai, etal. 𝜋0.5: avision-language-actionmodel with open-world generalization.arXiv preprint arXiv:2504.16054, 2025 on PithPith's reviewarXiv
show 80 more
  1. [21] Zhenyu Wu, Yuheng Zhou, Xiuwei Xu, Ziwei Wang, and Haibin Yan. Momanipvla: Transferring vision-language-action models for general mobile manipulation. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1714–1723. IEEE, 2025 on Pith
  2. [22] Xiaomi Robotics Team, Jun Guo, Piaopiao Jin, Jason Li, Peiyan Li, Yingyan Li, Futeng Liu, Wanli Peng, Optimus Qin, Yifei Su, et al. Xiaomi-robotics-1: Scaling vision-language-action models with over 100k hours of real-world trajectories.arXiv preprint arXiv:2607.15330, 2026 on PithPith's reviewarXiv
  3. [23] Wei Wu, Fangjing Wang, Fan Lu, He Sun, Shi Liu, Yunnan Wang, Yibin Yan, Yong Wang, Shuailei Ma, Xinyang Wang, et al. From foundation to application: Improving vla models in practice.arXiv preprint arXiv:2607.06403, 2026 on PithPith's reviewarXiv
  4. [24] Qixiu Li, Yu Deng, Yaobo Liang, Lin Luo, Lei Zhou, Chengtang Yao, Lingqi Zeng, Zhiyuan Feng, Huizhi Liang, Sicheng Xu, et al. Scalable vision-language-action model pretraining for robotic manipulation with real-life human activity videos.arXiv preprint arXiv:2510.21571, 2025 on PitharXiv
  5. [25] Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, et al. Qwen-vla: Unifying vision-language-action modeling across tasks, environments, and robot embodiments.arXiv preprint arXiv:2605.30280, 2026 on PithPith's reviewarXiv
  6. [26] Dongyoung Kim, Huiwon Jang, Myungkyu Koo, Suhyeok Jang, Taeyoung Kim, Beomjun Kim, Byungjun Yoon, Changsung Jang, Daewon Choi, Dongsu Han, et al. Rldx-1 technical report.arXiv preprint arXiv:2605.03269, 2026 on PithPith's reviewarXiv
  7. [27] Zheng Yang, Wenjie Zhang, Xiangyu Chen, Wenxuan Song, Xianpeng Wang, Yihang Kang, Jiawen Wen, WenChen, LujiaWang, RenjingXu, etal. Dreamtrajectory: Trajectory-guidedactiongeneration with world model alignment for mobile manipulation.arXiv preprint arXiv:2608.01381, 2026. 22 on PithPith's reviewarXiv
  8. [28] Zhe Li, Zhenzhe Zhang, Yangyang Wei, Wenjie Zhang, Xichen Yuan, Peiyuan Zhi, Gen Li, Xinying Guo, Fengjie Gao, Jianfei Yang, et al.𝜔-0: A latent predictive world action model for concurrent humanoid loco-manipulation.arXiv preprint arXiv:2608.06375, 2026 on PithPith's reviewarXiv
  9. [29] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first international conference on machine l... on Pith
  10. [30] MohitShridhar, LucasManuelli, andDieterFox. Perceiver-actor: Amulti-tasktransformerforrobotic manipulation. InConference on Robot Learning, pages 785–799. PMLR, 2023 on Pith
  11. [31] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations.arXiv preprint arXiv:2402.10885, 2024 on PithPith's reviewarXiv
  12. [32] Chengmeng Li, Junjie Wen, Yaxin Peng, Yan Peng, and Yichen Zhu. Pointvla: Injecting the 3d world into vision-language-action models.IEEE Robotics and Automation Letters, 11(3):2506–2513, 2026 on Pith
  13. [33] Lin Sun, Bin Xie, Yingfei Liu, Hao Shi, Tiancai Wang, and Jiale Cao. Geovla: Empowering 3d representations in vision-language-action models.arXiv preprint arXiv:2508.09071, 2025 on PithPith's reviewarXiv
  14. [34] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, et al. Spatialvla: Exploring spatial representations for visual-language-action model.arXiv preprint arXiv:2501.15830, 2025 on PithPith's reviewarXiv
  15. [35] Xiaoqi Li, Liang Heng, Jiaming Liu, Yan Shen, Chenyang Gu, Zhuoyang Liu, Hao Chen, Nuowei Han, Renrui Zhang, Hao Tang, et al. 3ds-vla: A 3d spatial-aware vision language action model for robust multi-task manipulation. In9th Annual Conference on Robot Learning, 2025 on Pith
  16. [36] Jiahui Zhang, Yurui Chen, Yueming Xu, Ze Huang, Yanpeng Zhou, Yu-Jie Yuan, Xinyue Cai, Guowei Huang, Xingyue Quan, Hang Xu, et al. 4d-vla: Spatiotemporal vision-language-action pretraining with cross-scene calibration.Advances in Neural Information Processing Systems, 38:33914... on Pith
  17. [37] JiamingLiu,QingpoWuwu,NuoweiHan,HaoChen,ZhuoyangLiu,FanFei,YueruJia,ChenyangGu, Yandong Guo, Boxin Shi, et al. Lift3d-vla: Lifting vla models to 3d geometry and dynamics-aware manipulation.arXiv preprint arXiv:2607.06564, 2026 on PithPith's reviewarXiv
  18. [38] Tao Lin, Gen Li, Yilei Zhong, Yanwen Zou, Yuxin Du, Jiting Liu, Encheng Gu, and Bo Zhao. Evo-0: Vision-language-action model with implicit spatial understanding.arXiv preprint arXiv:2507.00416, 2025 on PitharXiv
  19. [39] Yizhi Chen, Zhanxiang Cao, Xinyi Peng, Yixiao Zheng, Xiaxi Si, Yiheng Li, Liyun Yan, Keqi Zhu, Xueyun Chen, Shengcheng Fu, et al. Geoalign: Beyond semantics with state-guided spatial alignment in vla models.arXiv preprint arXiv:2606.03240, 2026 on PithPith's reviewarXiv
  20. [40] Mohan Liu, Zhihao Gu, Xuanyu Chen, Haitian Zhang, Kaimin Mao, Yan Wu, Wei-Yun Yau, and Lin Wang. Vistavla: Geometry-and semantic-aware 3d gaussian-grounded vla for robotic manipulation. arXiv preprint arXiv:2607.12356, 2026 on PithPith's reviewarXiv
  21. [41] Johan Bjorck, Fernando Castañeda, 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, 2025 on PithPith's reviewarXiv
  22. [42] 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... on PithPith's reviewarXiv
  23. [43] YandanYang,ShuangZeng,TongLin,XinyuanChang,DekangQi,JunjinXiao,HaoyunLiu,Ronghan Chen, Yuzhi Chen, Dongjie Huo, et al. Abot-m0: Vla foundation model for robotic manipulation with action manifold learning.arXiv preprint arXiv:2602.11236, 2026 on PithPith's reviewarXiv
  24. [44] Shangchen Miao, Ningya Feng, Jialong Wu, Ye Lin, Xu He, Dong Li, and Mingsheng Long. Jepa-vla: Video predictive embedding is needed for vla models.arXiv preprint arXiv:2602.11832, 2026 on PitharXiv
  25. [45] Lingchen Meng, Jianwei Yang, Rui Tian, Xiyang Dai, Zuxuan Wu, Jianfeng Gao, and Yu-Gang Jiang. Deepstack: Deeply stacking visual tokens is surprisingly simple and effective for lmms.Advances in Neural Information Processing Systems, 37:23464–23487, 2024 on Pith
  26. [46] ShuaiBai,YuxuanCai,RuizheChen,KeqinChen,XionghuiChen,ZesenCheng,LianghaoDeng,Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025 on PithPith's reviewarXiv
  27. [47] YulinLuo,HaoChen,ZhuangzheWu,BowenSui,JiamingLiu,ChenyangGu,ZhuoyangLiu,Qiuxuan Feng, Jiale Yu, Shuo Gu, et al. Look before acting: Enhancing vision foundation representations for vision-language-action models.arXiv preprint arXiv:2603.15618, 2026 on PitharXiv
  28. [48] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think.arXiv preprint arXiv:2410.06940, 2024 on PithPith's reviewarXiv
  29. [49] Jinghuan Shang, Karl Schmeckpeper, Brandon B May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling diverse vision foundation models for robot learning.arXiv preprint arXiv:2407.20179, 2024 on PithPith's reviewarXiv
  30. [50] Fuhao Li, Wenxuan Song, Han Zhao, Jingbo Wang, Pengxiang Ding, Donglin Wang, Long Zeng, and HaoangLi. Spatialforcing: Implicitspatialrepresentationalignmentforvision-language-actionmodel. InInternational Conference on Learning Representations, volume 2026, pages 132324–132345, 2026 on Pith
  31. [51] Minghao Guo, Meng Cao, Jiachen Tao, Rongtao Xu, Yan Yan, Xiaodan Liang, Ivan Laptev, and Xiaojun Chang. Glad: Geometric latent distillation for vision-language-action models.arXiv preprint arXiv:2512.09619, 2025 on PitharXiv
  32. [52] Guoheng Sun, Tingting Du, Kaixi Feng, Chenxiang Luo, Xingguo Ding, Zheyu Shen, Ziyao Wang, Yexiao He, and Ang Li. Rocket: Residual-oriented multi-layer alignment for spatially-aware vision-language-action models.arXiv preprint arXiv:2602.17951, 2026 on PitharXiv
  33. [53] Hao Wang, Xiaobao Wei, Jingyang He, Chengyu Bai, Chun-Kai Fan, Jiajun Cao, Jintao Chen, Ying Li, Shanyu Rong, Ming Lu, et al. Vega: Visual encoder grounding alignment for spatially-aware vision-language-action models.arXiv preprint arXiv:2605.10485, 2026 on PithPith's reviewarXiv
  34. [54] XingyuDing,YuzhongZhao,YangWu,ChaoyangZhao,ChunhaiZhao,YifanZhang,andJianCheng. Mind-vla: Instruction-awarespatialrepresentationalignmentforvision-language-actionmodels.arXiv preprint arXiv:2608.04633, 2026 on PithPith's reviewarXiv
  35. [55] Zonghe Liu, Shanyuan Jie, Xiaoquan Sun, Chen Cao, Zetian Xu, Zongsheng Liu, and Jiayu Chen. Sam3d-guided object-centric representation alignment for vision-language-action models.arXiv preprint arXiv:2607.25912, 2026 on PithPith's reviewarXiv
  36. [56] Yixuan Li, Yuhui Chen, Mingcai Zhou, Haoran Li, Zhengtao Zhang, and Dongbin Zhao. Qdepth-vla: Quantizeddepthpredictionasauxiliarysupervisionforvision-language-actionmodels.arXivpreprint arXiv:2510.14836, 2025. 24 on PithPith's reviewarXiv
  37. [57] Wenxuan Song, Ziyang Zhou, Han Zhao, Jiayi Chen, Pengxiang Ding, Haodong Yan, Yuxin Huang, Feilong Tang, Donglin Wang, and Haoang Li. Reconvla: Reconstructive vision-language-action model as effective robot perceiver. InProceedings of the AAAI Conference on Artificial Intellig... on Pith
  38. [58] Jiaxin Shi, Xidong Zhang, Fucai Zhu, Zhe Li, Siyu Zhu, and Weihao Yuan. 3dthinkvla: Endowing vision-language-actionmodelswithlatent3dpriorsvia3d-thinking-guidedco-training.arXivpreprint arXiv:2606.04436, 2026 on PithPith's reviewarXiv
  39. [59] Iok Tong Lei, Ying Jie Yap, Wei Huang, Qingchen Xie, Qianzhi Li, Yujie Zhang, Xiaolong Liu, and Zhidong Deng. Teaching tiny vla models where to look and how to move.arXiv e-prints, pages arXiv–2607, 2026 on Pith
  40. [60] Andrew Ting Yan Li, Zhuo Li, Zhelin Yang, Zhipeng Dong, Quentin Rouxel, and Fei Chen. Reasoning without inference cost: Latent semantic scaffolding for robot vla policies.arXiv preprint arXiv:2609.04893, 2026 on PithPith's reviewarXiv
  41. [61] Mengya Liu, Baoxiong Jia, Jiangyong Huang, Jingze Zhang, and Siyuan Huang. Lara: Latent action representation alignment for vision-language-action models.arXiv preprint arXiv:2606.07100, 2026 on PithPith's reviewarXiv
  42. [62] Xingyu Ding, Yuzhong Zhao, Chunhai Zhao, Yinghuan Shi, Chaoyang Zhao, and Yifan Zhang. Temporal forcing: 4d representation alignment for vision-language-action models.arXiv preprint arXiv:2608.30643, 2026 on PithPith's reviewarXiv
  43. [63] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. InInternational Conference on Learning Representations, volume 2024, pages 1064... on Pith
  44. [64] Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025 on PithPith's reviewarXiv
  45. [65] Jingwen Sun, Wenyao Zhang, Zekun Qi, Shaojie Ren, Zezhi Liu, Hanxin Zhu, Guangzhong Sun, Xin Jin, and Zhibo Chen. Vla-jepa: Enhancing vision-language-action model with latent world model. In European Conference on Computer Vision, pages 478–497. Springer, 2026 on Pith
  46. [66] Wenyao Zhang, Hongsi Liu, Zekun Qi, Yunnan Wang, Xinqiang Yu, Jiazhao Zhang, Runpei Dong, Jiawei He, He Wang, Zhizheng Zhang, et al. Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge.Advances in Neural Information Processing Systems, 38: 24195... on Pith
  47. [67] JisangHan,SeonghuJeon,JaewooJung,RenéZurbrügg,HonggyuAn,TifannyPortela,MarcoHutter, Marc Pollefeys, Seungryong Kim, and Sunghwan Hong. Geometric action model for robot policy learning.arXiv preprint arXiv:2606.17046, 2026 on PithPith's reviewarXiv
  48. [68] Lishan Yang, Wenxuan Song, Xi Wang, Pingyue Sheng, Zheng Fang, Ziyang Zhou, Junjie He, Haodong Yan, Jiayi Chen, Nan Sun, et al. 4d-wam: Infusing spatiotemporal awareness into world action models through trajectory fields.arXiv preprint arXiv:2608.08023, 2026 on PithPith's reviewarXiv
  49. [69] Davood Soleymanzadeh, Kaidi Zhang, Zhiyuan Zhang, Bihao Zhang, Xiao Liang, Yu She, and Minghui Zheng. Phr-vla: Planning horizon reasoning for vision-language-action models.arXiv preprint arXiv:2608.27609, 2026 on PithPith's reviewarXiv
  50. [70] Jianjun Zhang, Jian Zhu, Taiyi Su, Chong Ma, Zitai Huang, Yi Xu, and Hanli Wang. Learning 4d geometric priors for inference-efficient world action models.arXiv preprint arXiv:2607.05468, 2026. 25 on PithPith's reviewarXiv
  51. [71] Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. Vggt-𝜔.arXiv preprint arXiv:2605.15195, 2026 on PithPith's reviewarXiv
  52. [72] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022 on PithPith's reviewarXiv
  53. [73] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion- based generative models.Advances in neural information processing systems, 35:26565–26577, 2022 on Pith
  54. [74] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022 on PithPith's reviewarXiv
  55. [75] William Peebles and Saining Xie. Scalable diffusion models with transformers. In2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172–4182. IEEE, 2023 on Pith
  56. [76] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations.arXiv preprint arXiv:2403.03954, 2024 on PithPith's reviewarXiv
  57. [77] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. InInternational Conference on Learning Representations, volume 2025, pages 29982–30009, 2025 on Pith
  58. [78] Zifeng Gao, Guanxing Lu, Tianxing Chen, Wenxun Dai, Ziwei Wang, Chao Shang, Wenbo Ding, and YansongTang. Manicm: Real-time3ddiffusionpolicyviaconsistencymodelforroboticmanipulation. arXiv preprint arXiv:2406.01586, 2024 on PithPith's reviewarXiv
  59. [79] Lifeng Zhuo, Wendi Chen, Han Xue, Shirun Tang, Jun Lv, Cewu Lu, and Chuan Wen. Fa- rdp: A frequency-adaptive reactive diffusion policy for contact-rich manipulation.arXiv preprint arXiv:2607.28596, 2026 on PithPith's reviewarXiv
  60. [80] Prajwal Koirala and Mark Campbell. Vgfm: Expressive robot policies via dense value guidance in flow matching.arXiv preprint arXiv:2609.14261, 2026 on PithPith's reviewarXiv
  61. [81] TianhongLiandKaimingHe. Backtobasics: Letdenoisinggenerativemodelsdenoise. InProceedings oftheIEEE/CVFConferenceonComputerVisionandPatternRecognition,pages36115–36125,2026 on Pith
  62. [82] Yiyang Lu, Susie Lu, Qiao Sun, Hanhong Zhao, Zhicheng Jiang, Xianbang Wang, Tianhong Li, Zhengyang Geng, and Kaiming He. One-step latent-free image generation with pixel mean flows. arXiv preprint arXiv:2601.22158, 2026 on PithPith's reviewarXiv
  63. [83] Xianbang Wang, Hanhong Zhao, Yiyang Lu, Kangyang Zhou, Linrui Ma, and Kaiming He. Minit2i: Aminimalistbaselinefortext-to-imagegeneration,2026. URL https://peppaking8.github.io/ #/post/minit2i on Pith
  64. [84] Chaoyi Pan, Giridharan Anantharaman, Nai-Chieh Huang, Claire Jin, Daniel Pfrommer, Chenyang Yuan, Frank Permenter, Guannan Qu, Nicholas Boffi, Guanya Shi, et al. Much ado about noising: Dispelling the myths of generative robotic control. InInternational Conference on Learning ... on Pith
  65. [85] Josh Merel, Matthew Botvinick, and Greg Wayne. Hierarchical motor control in mammals and machines.Nature communications, 10(1):5489, 2019 on Pith
  66. [86] Arthur Allshire, Himanshu Gaurav Singh, Ritvik Singh, Adam Rashid, Hongsuk Choi, David McAllister, Justin Yu, Yiyuan Chen, Huang Huang, Pieter Abbeel, et al. Scalable behavior cloning with open data, training, and evaluation.arXiv preprint arXiv:2606.27375, 2026. 26 on PithPith's reviewarXiv
  67. [87] Shihan Wu, Xuecheng Liu, Shaoxuan Xie, Pengwei Wang, Xinghang Li, Bowen Yang, Zhe Li, Kai Zhu, Hongyu Wu, Yiheng Liu, et al. Robocoin: An open-sourced bimanual robotic data collection for integrated manipulation.arXiv preprint arXiv:2511.17441, 2025 on PithPith's reviewarXiv
  68. [88] ChengshuLi,RuohanZhang,JosiahWong,CemGokmen,SanjanaSrivastava,RobertoMartín-Martín, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. InConference on Robot Learni... on Pith
  69. [89] He Zhang, Lingzhu Xiang, Haitao Lin, Zeyu Huang, Minghui Wang, Dingyan Zhong, Yubo Dong, Yihao Wu, Yongming Rao, Dongsheng Zhang, et al. Hy-embodied-0.5-vla: From vision-language- action models to a real-world robot learning stack.arXiv preprint arXiv:2606.14409, 2026 on PithPith's reviewarXiv
  70. [90] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xuan Hu, Xu Huang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669, 2025 on PithPith's reviewarXiv
  71. [91] Yang Tian, Yuyin Yang, Yiman Xie, Zetao Cai, Xu Shi, Ning Gao, Hangxu Liu, Xuekun Jiang, Zherui Qiu, Feng Yuan, et al. Interndata-a1: Pioneering high-fidelity synthetic data for pre-training generalist policy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat... on Pith
  72. [92] RealSource. Realsource world: A large-scale real-world dual-arm manipulation dataset.https: //huggingface.co/datasets/RealSourceData/RealSource-World, 2025 on Pith
  73. [93] Tao Jiang, Tianyuan Yuan, Yicheng Liu, Chenhao Lu, Jianning Cui, Xiao Liu, Shuiqi Cheng, Jiyang Gao, Huazhe Xu, and Hang Zhao. Galaxea open-world dataset and g0 dual-system vla model.arXiv preprint arXiv:2509.00576, 2025 on PithPith's reviewarXiv
  74. [94] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12... on PithPith's reviewarXiv
  75. [95] AgiBotWorldTeam. Agibotworld2026. https://huggingface.co/datasets/agibot-world/ AgiBotWorld2026, 2026 on Pith
  76. [96] BitRobot, Unitree, and Hugging Face. Hiw-500: Humanoids in-the-wild dataset for robot learning. https://bitrobot-foundation.github.io/humanoids-in-the-wild-500-hours/, 2026 on Pith
  77. [97] Boyang Cai, Qiwei Liang, Jiawei Li, Shihang Weng, Zhaoxin Zhang, Tao Lin, Xiangyu Chen, Wenjie Zhang, Jiaqi Mao, Weisheng Xu, et al. Beyond viewpoint generalization: What multi- view demonstrations offer and how to synthesize them for robot manipulation?arXiv preprint arXiv:26... on PithPith's reviewarXiv
  78. [98] XPolicyLab Community, Tianxing Chen, Yue Chen, Tian Nian, Zijian Cai, Guangyu Chen, Wenwei Lin, Qiwei Liang, Peicheng Xiang, Kailun Su, et al. Xpolicylab: A unified standard and open ecosystem for robot policy evaluation and deployment.arXiv preprint arXiv:2608.09892, 2026 on PithPith's reviewarXiv
  79. [99] Ning Gao, Jinliang Zheng, Xing Gao, Haoxiang Ma, Hanqing Wang, Yukai Wang, Jiantong Chen, Zanxin Chen, Shujie Zhang, Mingda Jia, et al. Ebench: Elemental diagnosis of generalist mobile manipulation policies.arXiv preprint arXiv:2606.18239, 2026 on PithPith's reviewarXiv
  80. [100] Haoxiang Ma, Junhao Cai, Xiaoxu Xu, Hao Li, Yuyin Yang, Yang Tian, Jiafei Cao, Hongrui Zhu, Zherui Qiu, Yuqiang Yang, et al. Internvla-a1. 5: Unifying understanding, latent foresight, and action for compositional generalization.arXiv preprint arXiv:2607.04988, 2026. 27 on PithPith's reviewarXiv

The first 100 of 119 references are listed here. All of them are in this paper's citation record.

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