Pith. sign in

REVIEW 3 major objections 5 minor 52 references

The paper claims that one real demonstration, decomposed into a reusable contact skill and resynthesized across poses, scenes, and embodiments, can train open-world mobile manipulation policies.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 06:02 UTC pith:XI7WXTYS

load-bearing objection Strong data-engine paper with one real result and one overclaimed headline; the cross-environment generalization evidence is confounded by train/test leakage. the 3 major comments →

arxiv 2607.13154 v2 pith:XI7WXTYS submitted 2026-07-14 cs.RO

Worlds in One Demo: A Synthetic Data Engine for Learning Open-World Mobile Manipulation

Lingxiao Guo , Huanyu Li , Guanya Shi This is my paper
classification cs.RO
keywords mobile manipulationone-demonstration learningsynthetic data generationobject-relative replaycorrective state expansionGaussian splattingcross-embodiment generalizationscene generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

WANDA asks how much a single demonstration can teach. It reconstructs the background and the manipulated object from the demo, isolates the contact-rich interaction as an object-relative motion, and re-plans it into hundreds of new spatial configurations across both the reconstructed scene and new scenes generated from everyday photos. It deliberately injects off-nominal robot and object states so policies learn to recover from long-horizon drift, then renders photorealistic observations by compositing a Gaussian-splat background with a rendered robot and object. The paper reports that this one-demo pipeline trains policies reaching 54.8% average progress on five real-world mobile manipulation tasks across 16 environments, roughly matching or exceeding 20–50 teleoperated demonstrations in simulation, and transferring zero-shot to a different robot body with no demonstrations from that robot. If true, the human cost of mobile manipulation data collection collapses to one demonstration per task, and scene coverage becomes as easy as taking one photo.

Core claim

WANDA's central claim is that a long-horizon mobile manipulation task can be learned from one demonstration if the demonstration is decomposed into a reusable contact skill and a world model. The pipeline reconstructs a 3D background and object geometry from the demo's RGBD views, tracks the object's 6D pose, and treats the hand-object interaction as a segment to be replayed exactly in object-relative coordinates. It then resamples object placements, replans navigation and arm approach with whole-body inverse kinematics and sampling-based motion planning, and applies Corrective State Expansion to perturb robot and object states so the policy sees realistic drift. The result is a synthetic da

What carries the argument

Object-relative replay of contact-rich segments is the load-bearing mechanism: the end-effector motion is expressed relative to the manipulated object, so re-anchoring the object anywhere in the scene keeps the interaction kinematically consistent. Around that, whole-body inverse kinematics and sampling-based motion planning chain the replayed segment to new navigation and approach paths. Corrective State Expansion widens the distribution of object poses relative to the robot base and of robot start poses and arm states, teaching recovery from drift. Factorized rendering composites a Gaussian-splat background with a rendered robot and object foreground to produce photorealistic observations

Load-bearing premise

The load-bearing premise is that every successful manipulation transfers to new object poses by replaying the demonstrated hand-object motion in object-relative coordinates, without simulating forces, friction, or deformation; if contact skills depend on those, the synthesized trajectories will not be executable.

What would settle it

Give WANDA one demonstration of a force-sensitive or deformable task—such as folding a towel or inserting a plug into a socket—and train a policy on the generated set. If, unlike the reported rigid-object tasks, the policy cannot complete the task at any novel object pose, the kinematic-replay assumption is falsified. A stricter check: for a rigid task, simulate the generated trajectories with physics and inspect whether the replayed contact ever loses grasp, penetrates, or deforms unrealistically under the object perturbations.

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

If this is right

  • If the central claim is correct, per-task data collection drops from hundreds of teleoperated demonstrations to one recorded demonstration, with annotation costing only minutes.
  • Scene generalization becomes photo-driven: a single picture of a new room can generate training data that transfers to the real room without taking the robot there.
  • Cross-embodiment transfer becomes a natural byproduct for non-prehensile tasks, since the reconstructed world and object-relative contact replay are embodiment-agnostic.
  • Dataset scaling becomes a compute problem rather than a human-collection problem, with both more generated trajectories and more generated scenes shown to improve task performance.
  • Long-horizon robustness can be learned from deliberately injected drift instead of waiting for naturally occurring failures, which may transfer to other imitation-learning pipelines.

Where Pith is reading between the lines

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

  • The same object-relative replay principle could extend to multi-object contact graphs, such as stacking or tool use, by treating each contact pair as a node and reconciling simultaneous constraints.
  • The rigid-object ceiling suggests a crisp test of the method's own boundary: soft or deformable tasks will likely fail under pure kinematic replay, pointing to dynamics or deformation modeling as the natural next component.
  • If scene generalization holds broadly, the open-world mobile manipulation bottleneck may shift from data collection to reconstruction and planning reliability, since a single photo of a scene suffices for training data.
  • The 5–30 minutes of annotation per demonstration is a residual human cost that appears automatable, because segmentation and region selection are already driven by vision-language models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces WANDA, a data-generation engine that converts a single real RGBD demonstration into many synthetic mobile-manipulation trajectories. The pipeline reconstructs the background as 2D Gaussian splats and the foreground object as mesh plus tracked 6D pose, then rearranges contact-rich segments into new spatial configurations via whole-body IK/RRT-Connect planning, applies Corrective State Expansion (CSE) to broaden robot and object state coverage, and renders photo-realistic observations by compositing Isaac Sim foregrounds with Gaussian-splat backgrounds. It also uses Marble-generated 3D worlds from single photos for scene diversity. The authors report simulation results on Bigym and BEHAVIOR Challenge, real-world results on five long-horizon tasks, a scalability study, and zero-shot cross-embodiment deployment on Linearbot, claiming ~50x data efficiency and broad generalization from one demonstration.

Significance. If the core claims hold, WANDA would be a meaningful step toward drastically reducing human data-collection effort for open-world mobile manipulation. The paper has several genuine strengths: a detailed, mostly self-contained pipeline description (Appendices A-H), explicit per-subtask counts alongside aggregate progress scores, a clear CSE ablation, per-task data-generation cost breakdowns, and an honest limitations section. The simulation evidence that one WANDA-generated dataset can compete with or beat 20-50 teleoperated demos is encouraging. However, the headline “cross-environment generalization” claim is not currently supported because the evaluation uses Marble worlds generated from photos of the same scenes used for testing, and the real-world results rest on only 10 rollouts per task with no uncertainty quantification. These issues are fixable but require either new experiments or substantially softened claims.

major comments (3)
  1. [Section 4.3 / Figure 7(b)] The claim of “cross-environment generalization” is confounded by train/test leakage. The paper states that for each evaluation scene, a photo is taken and fed into Marble, and the policy is trained on data synthesized from that photo-generated world and then evaluated in the same real scene. This tests same-scene synthetic-to-real transfer, not generalization to a scene whose photo was never used for training. The source-only baseline does not remove the confound; it only shows that source-scene data alone does not transfer. A valid test requires held-out scenes whose photos are never used to generate training worlds. The abstract and Section 4.3 should either provide such an experiment or rephrase the claim as “photo-conditioned scene adaptation” rather than cross-environment generalization.
  2. [Table 2 / Figure 6] The real-world results are reported as means over 10 rollouts per task with no error bars, confidence intervals, or statistical tests. For the central CSE ablation, the gap between 54.8% and 15.7% is large, but raw subtask counts such as 6/10, 5/10, or 3/10 show that individual task means could shift substantially with a few outcomes changed. The paper's headline “broad spatial and long-horizon robustness” would be much better supported by per-task Wilson intervals, bootstrap CIs, or additional rollouts. Without this, the real-world quantitative claims are suggestive rather than statistically established.
  3. [Section 4.4 / Table 4] The zero-shot cross-embodiment claim is based on a single task (Drop Trash) with 10 rollouts on Linearbot and no comparison against training directly on Linearbot data or against a baseline that uses the same amount of Linearbot teleoperation data. The statement “first zero-shot cross-embodiment generalization capability in mobile manipulation” exceeds what one task, one robot pair, and 10 rollouts can support. I recommend either adding more tasks/robots, adding statistical precision, or tempering the claim to “a demonstration of feasibility on one task.”
minor comments (5)
  1. [Section 5 / Eq. (B.3)] The paper correctly acknowledges that only rigid and articulated objects are modeled, while soft objects and fluids are handled by replaying the demonstrated motion. Since POUR involves pouring, it would help readers to state explicitly whether POUR relies on this replay mode and whether Object State Expansion perturbs the teapot in ways that could invalidate the replayed pouring motion. This is a clarity point, not a fatal flaw, because the limitation is already disclosed.
  2. [Table 2] The note “The Utensil ablation retains its legacy two-subtask score” is unexplained. If the Utensil task definition changed between the full WANDA and ablation evaluations, the comparison may not be apples-to-apples. Please specify the task versions and whether per-subtask definitions are identical.
  3. [Figure 7(a)] The x-axis “Data proportion” is ambiguous. It appears to denote the fraction of the full generated dataset used for training, but the reader must infer this from the text. Please label clearly and state what 0%, 20%, 75%, and 100% correspond to in terms of number of trajectories.
  4. [Appendix B.1] Equation (B.1) introduces many cost weights and tolerances, but no default values or selection procedure are given. Because the method's behavior depends on these constants, providing a table of all configuration constants (e.g., R_t, R_psi, alpha, epsilon_p, epsilon_R, w_*) would substantially improve reproducibility.
  5. [Figure 6 / Appendix C.2] The conditional subtask metric is appropriate, but drawing 0/0 as a short stub is visually confusing because it can look like a legitimate low success rate. Consider using a distinct marker (e.g., “N/A”) or adding an explicit note in the figure caption.

Circularity Check

1 steps flagged

Cross-environment generalization claim is confounded: training worlds are generated from photos of the evaluation scenes themselves.

specific steps
  1. fitted input called prediction [Section 4.3, 'Scaling generated scenes for open-world generalization' (Figure 7(b)); abstract 'cross-environment generalization']
    "For each scene, we take a single photo and feed it into Marble to generate a corresponding 3D world. We train the policy on data synthesized from both source scenes and these photo-generated worlds, and evaluate it in the same real-world scenes where the photos were taken. Although the generated worlds are derived from these environments, they still contain substantial visual and geometric discrepancies from reality."

    The cross-environment generalization 'prediction' is not a prediction about unseen scenes: for each evaluation scene, a photo of that very scene is converted by Marble into a training world. The policy is therefore trained on data whose scene identity, room layout, tables, and object placements come from the exact scene in which it is later evaluated. The source-only near-zero baseline only shows that scene-specific generated data helps in that same scene; it does not establish generalization to a scene whose photo was never used. The abstract's 'cross-environment generalization' claim is thus supported only by a train-on-the-test-scene construction rather than by a held-out scene experiment.

full rationale

WANDA's core data-generation pipeline is not circular in the mathematical sense: Eq. (B.3) is the standard object-relative replay definition, and the long-horizon and spatial-generalization claims are backed by physical real-robot rollouts with randomized object placements, so those experiments have independent content. The cross-embodiment experiment also has content: the policy is trained on synthetic Linearbot data and deployed zero-shot on the real Linearbot, with no real Linearbot demonstrations. The one load-bearing circular step is the cross-environment evaluation in Section 4.3: a photo of each evaluation scene is used to generate a Marble training world, and the policy is then evaluated in that same real scene. The paper's own sentence 'the generated worlds are derived from these environments' concedes the leak. The rebuttal that renderings still differ visually from reality addresses sim-to-real gap, not scene novelty: the scene identity and geometry are shared between training and test, so the reported 'novel scene' progress is forced by construction once the test scene's photo enters the training pipeline. Section 5's deformable-object limitation is real but is a scope limitation, not a circularity. Because one of the abstract's headline claims—cross-environment generalization—reduces by construction while the other claims retain independent experimental grounding, the appropriate score is 6 (partial circularity).

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities or conserved quantities. It postulates a method (CSE) and relies on a set of domain assumptions about the transferability of kinematic contact replay, reconstruction fidelity, and the validity of photo-derived test scenes. The free parameters are engineering tuning choices for the data engine, none of which are reported numerically or ablated for sensitivity.

free parameters (4)
  • CSE perturbation ranges (R_t, R_psi, alpha)
    Object State Expansion draws translation/yaw from configured ranges and Robot State Expansion samples a start in a cone of half-angle alpha (Eqs. B.4, B.8). These hand-chosen ranges directly determine the state diversity of the generated data; no values or sensitivity analysis are reported in the paper.
  • IK and planning tolerances (epsilon_p, epsilon_R) and cost weights (w_rho, w_c, w_f, w_xy, w_psi)
    Accepted trajectories are filtered by position/rotation tolerances (Eq. B.7) and navigation candidates are ranked by a weighted cost (Eq. B.1). These constants control the quality and diversity of synthesized trajectories and are chosen by the authors without reported values or ablations.
  • Number of generated demonstrations per task (263-740, plus 400 for Linearbot) = 263-740 per real task; 400 cross-embodiment
    Dataset sizes per task were selected by the authors; scaling experiments show performance improves with more data, so these choices affect the reported progress scores.
  • VLM scale estimation for Marble scenes
    Since Marble-generated scenes lack metric scale, the VLM 'leverages commonsense knowledge to infer plausible entity heights' (Section 3.1). This is an ad hoc calibration that affects all planning and rendering in generated worlds and has no ground-truth check.
axioms (5)
  • domain assumption Rigid and articulated objects only; soft objects replayed without deformation modeling
    Section 5 states WANDA handles only rigid and articulated objects and 'replays the demonstrated motion rather than modeling the underlying deformation' for cloth/fluids. The central spatial-generalization claim depends on this restriction.
  • domain assumption Object-relative kinematic replay preserves the contact-rich interaction (Eq. B.3)
    The planner transfers source EE poses via T_Ek_W(i) = T_o_W (T_o,src_W(i))^-1 T_Ek,src_W(i), assuming the task-relevant interaction is fully encoded in 6D poses and can be replayed without dynamics or force feedback. This is the load-bearing premise for spatial generalization.
  • domain assumption MAtCha, BundleSDF, and Marble produce sufficiently accurate geometry, pose, texture, and scale from a single demonstration's observations
    The pipeline treats these external reconstruction/generation tools as black boxes. If any stage produces poor geometry or scale, all downstream planning and rendered observations inherit the error. The paper relies on these tools without independent validation beyond qualitative figures.
  • domain assumption VLM commonsense provides valid randomization regions and object heights
    Gemini 3.1 Pro is prompted to identify initialization regions from rendered orbit views and infer metric scale for Marble scenes (Section 3.1). The quality of generated spatial configurations depends on the VLM's correctness.
  • domain assumption Testing in real scenes whose photos were used to generate the training worlds is a valid test of cross-environment generalization
    Section 4.3 evaluates policies trained on Marble worlds in 'the same real-world scenes where the photos were taken.' The paper asserts the generated worlds differ enough from reality to require generalization, but the test remains partially in-distribution because the test environments directly inform the synthetic training data.

pith-pipeline@v1.3.0-alltime-deepseek · 19540 in / 12522 out tokens · 109918 ms · 2026-08-02T06:02:06.080147+00:00 · methodology

0 comments
read the original abstract

Learning open-world mobile manipulation policies requires vast data to achieve spatial generalization, long-horizon robustness, and scene generalization. Current prevailing data collection paradigms, teleoperation and UMI, demand prohibitive human effort and cost at scale. To scale beyond the limits of manual data collection, we seek to maximize the value of each human demonstration by scalable data generation. To this end, we introduce WANDA: learning open-World mobile mANipulation from one demonstration via a synthetic DAta engine. WANDA first reconstructs background Gaussian splats and robot-object interaction trajectories from source RGBD observations, as a world substrate for later planning and rendering. It then rearranges contact-rich robot-object interaction segments into extensive spatial configurations, utilizing whole-body motion planning to chain them into new trajectories. To enhance long-horizon robustness, it applies Corrective State Expansion to increase the robot and object state diversity at different stages of mobile manipulation. To unlock cross-environment generalization, trajectories are synthesized on diverse generated 3D worlds from everyday photos. Furthermore, we synthesize photo-realistic observations by compositing rendered robot and object meshes with Gaussian splatting backgrounds. We evaluate our approach on extensive simulation and real-world tasks in various scenes. Experiments show that policies trained with WANDA achieve long-horizon robustness, broad spatial generalization and cross-environment generalization from one real demonstration. Moreover, WANDA naturally supports cross-embodiment data generation, validated by zero-shot deployment on another mobile manipulator with a distinct morphology.

Figures

Figures reproduced from arXiv: 2607.13154 by Guanya Shi, Huanyu Li, Lingxiao Guo.

Figure 1
Figure 1. Figure 1: WANDA enables mobile manipulation policies achieve extreme spatial generalization, long hori￾zon robustness, cross-environment generalization and cross-embodiment generalization from only one human demonstration. This is achieved by the proposed data engine that synthesizes diverse trajectories across reconstructed worlds and generated worlds on different robot embodiments. Correspondence: {lingxiag, huany… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline overview of WANDA. From a single real demonstration, WANDA reconstructs and gener￾ates 3D worlds, rearranges contact-rich robot-object interaction segments into diverse spatial configurations, chains them into new trajectories with whole-body motion planning, applies Corrective State Expansion to increase robot and object state coverage, and produces photo-realistic observations by factorized rend… view at source ↗
Figure 3
Figure 3. Figure 3: Sampled initial configuration results on reconstructed and generated 3D scenes. Scene and robot meshes are shown in Blender. Red regions indicate the object initial pose sampling regions. source observation. This acts as a world substrate for later planning and rendering. For background reconstruction, we first subsample a set of diverse mobile camera views from the source demonstra￾tions. Then we inpaint … view at source ↗
Figure 4
Figure 4. Figure 4: Real-world tasks. Five long-horizon mobile-manipulation tasks, each learned from a single real demonstration: LUNCHBOX, UTENSIL, DROPTRASH, POUR, and FRIDGE [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Demonstration of simulation tasks and subtasks from Bigym [Chernyadev et al., 2024] and B1K [BEHAVIOR Team, 2025] [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-subtask success rate on the real-world tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scalability results. (a) Performance versus generated-data proportion. (b) Per-task progress scores in three novel scenes; each task–scene pair compares WANDA with the 0% source-only baseline. • Pour: The robot needs to navigate to the teapot randomized in the large kitchen table, take the teapot to the cup and pour into it. 740 trajectories are generated for the task. • Fridge: The robot is required to ta… view at source ↗
Figure 8
Figure 8. Figure 8: Per-subtask success rate across novel scenes. WANDA [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cross-embodiment data generation from Agibot G1 to Linearbot [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Corrective State Expansion. 5 Limitations and Future Work WANDA has two main limitations. First, it handles only rigid and articulated objects. Reconstruction tracks the 6D poses of rigid parts and replays them, so soft and deformable objects, such as cloth and fluids, cannot be reconstructed or generated; in such cases WANDA replays the demonstrated motion rather than modeling the underlying deformation.… view at source ↗
Figure 11
Figure 11. Figure 11: Demonstration post-process interface used to annotate contact-rich and navigation segments. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Background reconstruction results (textured mesh and geometry mesh) visualized in Blender. From left to right: Pour, Drop Trash, Fridge, Utensil. A.2 Background Reconstruction Given inpainted background images, we utilize a pose-free sparse-view 2D Gaussian splatting method, MAtCha [Guédon et al., 2025] to reconstruct both splats and geometry meshes, and predict camera poses in the world frame. To align t… view at source ↗
Figure 13
Figure 13. Figure 13: Foreground reconstruction: BundleSDF vs. BundleSDF with generative completion. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Additional foreground reconstruction results [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Initial configuration results on reconstructed and generated 3D scenes. A.4 VLM-guided Initial Configuration We provide the detailed prompt used in Appendix H.2. Here we show the results of initial configura￾tion on both reconstructed 3D scenes and generated 3D scenes. The VLM successfully recognize the reasonable region for randomizing the objects, and place the objects in extensive spaces to enhance the… view at source ↗
Figure 16
Figure 16. Figure 16: Representative frames from the real-scene trajectories. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Representative frames from the world-lab trajectories. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Representative frames from the world-lab trajectories (continued). 28 [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: shows the generated visual observations for the cross-embodiment data generation from Agibot G1 to Linearbot on Drop Trash, described in Section 4.4. Despite a large kinematic gap (different DoFs) and a large visual gap (different camera extrinsics and intrinsics) between the two robots, WANDA generates reasonable visual observations [PITH_FULL_IMAGE:figures/full_fig_p029_19.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 21 linked inside Pith

  1. [1]

    Proceedings of The 7th Conference on Robot Learning , pages =

    MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations , author =. Proceedings of The 7th Conference on Robot Learning , pages =. 2023 , volume =

    2023

  2. [2]

    arXiv preprint arXiv:2502.16932 , year =

    DemoGen: Synthetic Demonstration Generation for Data-Efficient Visuomotor Policy Learning , author =. arXiv preprint arXiv:2502.16932 , year =

    Pith/arXiv arXiv

  3. [3]

    2025 , eprint =

    MoMaGen: Generating Demonstrations under Soft and Hard Constraints for Multi-Step Bimanual Mobile Manipulation , author =. 2025 , eprint =

    2025

  4. [4]

    arXiv preprint arXiv:2505.09601 , year =

    Real2Render2Real: Scaling Robot Data Without Dynamics Simulation or Robot Hardware , author =. arXiv preprint arXiv:2505.09601 , year =

    Pith/arXiv arXiv

  5. [5]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    BundleSDF: Neural 6-DoF Tracking and 3D Reconstruction of Unknown Objects , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  6. [6]

    arXiv preprint arXiv:2407.07788 , year =

    BiGym: A Demo-Driven Mobile Bi-Manual Manipulation Benchmark , author =. arXiv preprint arXiv:2407.07788 , year =

    Pith/arXiv arXiv

  7. [7]

    2025 , howpublished =

    2025

  8. [8]

    arXiv preprint arXiv:2504.16054 , year =

    _. arXiv preprint arXiv:2504.16054 , year =

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2604.15483 , year =

    _. arXiv preprint arXiv:2604.15483 , year =

    Pith/arXiv arXiv

  10. [10]

    2026 , howpublished =

    Marble Documentation , author =. 2026 , howpublished =

    2026

  11. [11]

    The International Journal of Robotics Research , year =

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author =. The International Journal of Robotics Research , year =

  12. [12]

    arXiv preprint arXiv:2510.20774 , year =

    FieldGen: From Teleoperated Pre-Manipulation Trajectories to Field-Guided Data Generation , author =. arXiv preprint arXiv:2510.20774 , year =

    arXiv

  13. [13]

    arXiv preprint arXiv:2505.23692 , year =

    Mobi- : Mobilizing Your Robot Learning Policy , author =. arXiv preprint arXiv:2505.23692 , year =

    arXiv

  14. [14]

    ICRA 2025 Workshop: Human-Centered Robot Learning in the Era of Big Data and Large Models , year =

    DynaMem: Online Dynamic Spatio-Semantic Memory for Open World Mobile Manipulation , author =. ICRA 2025 Workshop: Human-Centered Robot Learning in the Era of Big Data and Large Models , year =

    2025

  15. [15]

    arXiv preprint arXiv:2503.05652 , year =

    BEHAVIOR Robot Suite: Streamlining Real-World Whole-Body Manipulation for Everyday Household Activities , author =. arXiv preprint arXiv:2503.05652 , year =

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2401.02117 , year =

    Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation , author =. arXiv preprint arXiv:2401.02117 , year =

    Pith/arXiv arXiv

  17. [17]

    8th Annual Conference on Robot Learning , year =

    SkillMimicGen: Automated Demonstration Generation for Efficient Skill Learning and Deployment , author =. 8th Annual Conference on Robot Learning , year =

  18. [18]

    Conference on Robot Learning (CoRL) , year =

    One Demo is Worth a Thousand Trajectories: Action-View Augmentation for Visuomotor Policies , author =. Conference on Robot Learning (CoRL) , year =

  19. [19]

    arXiv preprint arXiv:2605.02881 , year=

    MolmoAct2: Action Reasoning Models for Real-world Deployment , author=. arXiv preprint arXiv:2605.02881 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2406.09246 , year=

    Openvla: An open-source vision-language-action model , author=. arXiv preprint arXiv:2406.09246 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2602.15922 , year=

    World action models are zero-shot policies , author=. arXiv preprint arXiv:2602.15922 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2407.01512 , year=

    Open-television: Teleoperation with immersive active visual feedback , author=. arXiv preprint arXiv:2407.01512 , year=

    Pith/arXiv arXiv

  23. [23]

    2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Learning human-to-humanoid real-time whole-body teleoperation , author=. 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2024 , organization=

    2024

  24. [24]

    arXiv preprint arXiv:2406.10454 , year=

    Humanplus: Humanoid shadowing and imitation from humans , author=. arXiv preprint arXiv:2406.10454 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2402.10329 , year=

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots , author=. arXiv preprint arXiv:2402.10329 , year=

    Pith/arXiv arXiv

  26. [26]

    Conference on Robot Learning , pages=

    FastUMI: A scalable and hardware-independent universal manipulation interface with dataset , author=. Conference on Robot Learning , pages=. 2025 , organization=

    2025

  27. [27]

    arXiv preprint arXiv:2603.03243 , year=

    Hommi: Learning whole-body mobile manipulation from human demonstrations , author=. arXiv preprint arXiv:2603.03243 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2602.03310 , year=

    RDT2: Exploring the Scaling Limit of UMI Data Towards Zero-Shot Cross-Embodiment Generalization , author=. arXiv preprint arXiv:2602.03310 , year=

    arXiv

  29. [29]

    arXiv preprint arXiv:2505.12705 , year=

    Dreamgen: Unlocking generalization in robot learning through video world models , author=. arXiv preprint arXiv:2505.12705 , year=

    Pith/arXiv arXiv

  30. [30]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  31. [31]

    ICRA 2026 Workshop on Synthetic Data for Robot Learning , year=

    HumanoidMimicGen: Data Generation for Loco-Manipulation via Whole-Body Planning and Adaptation , author=. ICRA 2026 Workshop on Synthetic Data for Robot Learning , year=

    2026

  32. [32]

    arXiv preprint arXiv:2510.08547 , year=

    R2RGEN: Real-to-Real 3D Data Generation for Spatially Generalized Manipulation , author=. arXiv preprint arXiv:2510.08547 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2512.19402 , year=

    Real2Edit2Real: Generating Robotic Demonstrations via a 3D Control Interface , author=. arXiv preprint arXiv:2512.19402 , year=

    arXiv

  34. [34]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Matcha gaussians: Atlas of charts for high-quality geometry and photorealism from sparse views , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  35. [35]

    arXiv preprint arXiv:2602.06949 , year=

    DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos , author=. arXiv preprint arXiv:2602.06949 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2311.15127 , year=

    Stable video diffusion: Scaling latent video diffusion models to large datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2501.09747 , year=

    Fast: Efficient action tokenization for vision-language-action models , author=. arXiv preprint arXiv:2501.09747 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2403.09631 , year=

    3d-vla: A 3d vision-language-action generative world model , author=. arXiv preprint arXiv:2403.09631 , year=

    Pith/arXiv arXiv

  39. [39]

    ACM SIGGRAPH 2024 conference papers , pages=

    2d gaussian splatting for geometrically accurate radiance fields , author=. ACM SIGGRAPH 2024 conference papers , pages=

    2024

  40. [40]

    Proceedings 2000 ICRA

    RRT-connect: An efficient approach to single-query path planning , author=. Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065) , volume=. 2000 , organization=

    2000

  41. [41]

    ACT-1: A Robot Foundation Model Trained on Zero Robot Data , year =

  42. [42]

    arXiv preprint arXiv:2604.07607 , year=

    Egoverse: An egocentric human dataset for robot learning from around the world , author=. arXiv preprint arXiv:2604.07607 , year=

    Pith/arXiv arXiv

  43. [43]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Egomimic: Scaling imitation learning via egocentric video , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  44. [44]

    arXiv preprint arXiv:2602.16710 , year=

    Egoscale: Scaling dexterous manipulation with diverse egocentric human data , author=. arXiv preprint arXiv:2602.16710 , year=

    arXiv

  45. [45]

    European conference on computer vision , pages=

    Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  46. [46]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Sam 3d: 3dfy anything in images , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  47. [47]

    Advances in neural information processing systems , volume=

    Implicit neural representations with periodic activation functions , author=. Advances in neural information processing systems , volume=

  48. [48]

    Communications of the ACM , volume=

    Nerf: Representing scenes as neural radiance fields for view synthesis , author=. Communications of the ACM , volume=. 2021 , publisher=

    2021

  49. [49]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  50. [50]

    arXiv preprint arXiv:2511.07416 , year=

    Robot Learning from a Physical World Model , author=. arXiv preprint arXiv:2511.07416 , year=

    arXiv

  51. [51]

    arXiv preprint arXiv:2512.16881 , year=

    Polaris: Scalable real-to-sim evaluations for generalist robot policies , author=. arXiv preprint arXiv:2512.16881 , year=

    arXiv

  52. [52]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    4d gaussian splatting for real-time dynamic scene rendering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=