Pith. sign in

REVIEW 3 major objections 8 minor 35 references

FloAff-Kitchen: Bridging Navigation and Manipulation via Canonical and Progressive Floor Affordance Learning

T0 review · 3 major / 8 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Where a mobile robot should stand for manipulation is a local geometry problem that becomes learnable once scenes are canonicalized and skills are taught in stages.

desk verdict Solid MoMa engineering paper: canonical local geometry plus progressive priors beat prior FloAff baselines, with a useful multi-view kitchen benchmark—and the near-Oracle SR claim is mostly map fidelity under the same labeling policy. read the letter →

arxiv 2607.24207 v1 pith:KO3Q5RGM submitted 2026-07-27 cs.RO

Ping Zhong , Manling Teng , Tao Wu , Bolei Chen , Jiazhi Xia , Jianxin Wang This is my paper
classification cs.RO
keywords mobilemanipulationflooraffordancebaseplacementcanonicalrepresentationprogressivemulti-tasklearningegocentricRGB-Dkitchenbenchmarkviewpointrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Mobile manipulation fails when the robot only navigates to a feasible spot and not to a standing pose that makes the next grasp or door motion succeed. This paper argues that those standing regions—Floor Affordance—are fixed mainly by local interaction geometry around the target, not by the whole kitchen, and that shared placement rules (reach, collision, visibility) can be learned first from pick-and-place and then adapted to doors and drawers. It builds a canonical floor representation that crops to the target and rotates objects into a shared facing frame, then trains affordance maps progressively so common priors are not drowned by skill-specific constraints. A new multi-view kitchen benchmark stresses style, layout, and side-view changes. On existing and new tests the approach predicts better affordance maps and yields higher downstream success rates, approaching an oracle base pose.

What carries the argument

Canonical Floor Affordance Representation (CFAR) plus Progressive Floor Affordance Learning (PFAL): CFAR maps the scene to a target-centered, yaw-aligned local point cloud; PFAL pretrains on pick-and-place then adapts to door/drawer tasks while a distillation loss keeps the shared prior.

What would settle it

Hold the same kitchen scenes and CFAR pipeline, but relabel FloAff with a different manipulation policy (or real-robot rollouts) and check whether predicted maps and multi-task success still beat the same baselines by a similar margin on Styles and Layouts side- and front-view splits.

Watch

Extended reading notes

Core claim

Floor Affordance prediction is target-conditioned local spatial reasoning: if egocentric scenes are turned into a canonical interaction geometry (crop around the target, align orientation) and transferable placement priors are first learned on pick-and-place then adapted to heterogeneous skills with a shared-feature distillation term, the same model predicts manipulation-ready base maps more accurately across kitchen styles, layouts, and viewpoints than joint multi-task baselines.

Load-bearing premise

Ground-truth standing regions are defined by whether one fixed pretrained manipulation policy succeeds from collision-filtered poses in simulation, so the maps are policy- and simulator-specific rather than pure physical affordance.

Editorial extensions

If this is right

  • Base placement for MoMa should optimize predicted manipulation success maps, not only free-space navigation goals.
  • Cropping and orientation alignment alone improve single-task FloAff-grounded success, especially on hinge-sensitive door tasks.
  • Training pick-and-place first, then doors/drawers with feature distillation, raises cross-skill success without full joint entanglement.
  • Multi-view kitchen benchmarks with shared rollout labels can measure viewpoint robustness separately from style and layout shift.
  • Predicted FloAff can sit in front of existing manipulation policies and approach oracle standing performance in sim kitchens.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If labels stay policy-tied, FloAff models may need periodic relabeling whenever the downstream arm policy changes, or a multi-policy labeling set to approximate policy-agnostic maps.
  • The same canonicalize-then-curriculum pattern may transfer to non-kitchen MoMa (tables, shelves) if a weak-constraint foundation skill still exists.
  • Test-time yaw errors from the orientation net are a hidden failure mode on side views; measuring map quality conditional on yaw error would isolate that bottleneck.
  • Near-oracle sim success suggests remaining MoMa gaps may shift from where to stand toward when the arm policy itself fails under residual occlusion.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. The paper addresses Floor Affordance (FloAff) prediction for mobile manipulation: predicting, from egocentric RGB-D plus proprioception, a ground-plane map of base placements that enable successful task execution. It makes three contributions. First, CFAR canonicalizes observations by cropping a target-centered local point cloud (radius r) and rotating it into a target-aligned frame using a yaw angle that is simulator-provided at training time and predicted by a small AxisNet module at inference. Second, PFAL trains a FloAff predictor progressively — pick-and-place first, then joint adaptation to hinge/sliding tasks — with a feature-distillation loss intended to preserve shared priors. Third, the authors build FloAff-Kitchen, a RoboCasa-based benchmark with style and layout splits and both front- and side-view observations, labeled by rollout success of a pretrained task policy at collision-filtered candidate base poses. Experiments on MoMa-Kitchen and the two new benchmarks show consistent improvements over NavAff, N2M, C2F-Exp, and point-cloud backbones on map metrics (RMSE/logMSE/PCC/SIM) and on affordance-grounded success rate, with ablations isolating crop, alignment, Soft Dice loss, and feature distillation.

Significance. If the results hold, the paper contributes (i) a well-motivated canonicalization pipeline whose two components (cropping, alignment) are each ablated and shown to matter, (ii) a progressive training scheme with a distillation objective that is also ablated, and (iii) the first multi-view FloAff benchmark with explicit front/side-view protocols, which the authors commit to releasing with code and evaluation protocols — a genuine community asset, since viewpoint robustness in base-placement prediction is essentially unstudied. The reported gains are large (MoMa-Kitchen RMSE 0.073 vs 0.147 for the prior SOTA) and consistent across three settings, two task granularities, and per-task breakdowns. The main caveats on significance are that all evidence is simulated (RoboCasa/MoMa-Kitchen), the learned affordance is conditioned on one specific policy/simulator labeling stack, and test scenes number one per task, so the strength of the generalization claims should be read accordingly.

major comments (3)
  1. [Rollout-based FloAff Annotation, Eq. (2)/(16); Table 2] Eq. (2)/(16) and Table 2: the GT FloAff labels are binary rollout successes of a fixed pretrained policy stack (Diffusion Policy, Chi et al. 2025) at collision-filtered candidate poses, and the downstream SR in Table 2 is measured by executing those same policies from poses selected on the predicted map. Consequently, any predictor that matches the GT map well is near-Oracle in SR essentially by construction; Table 2 is largely a re-expression of Table 1's map fidelity rather than independent evidence that better maps 'translate into improved manipulation performance beyond affordance prediction accuracy' (§Comparative Studies). More importantly, the learned quantity is 'success region of this particular policy checkpoint, simulator, and collision filter,' not a policy-agnostic physical affordance, yet the introduction motivates FloAff as complementary to and integrable with 'different m
  2. [Implementation Details; Parametric Studies, Tables 6–7] The crop radii used for the headline results (r = 1.30 m Styles, 1.20 m Layouts) are exactly the best rows of Tables 6–7, which are computed on the respective test benchmarks. This is hyperparameter selection on the test set. On Styles the sensitivity is small (PCC 0.846–0.863 across radii), so the main margins almost certainly survive; on Layouts the sensitivity is larger (PCC 0.662 at r=1.10 vs 0.738 at r=1.20). Given that each task has only one unseen test scene, per-benchmark test-set tuning weakens the comparison protocol. Please select r on a held-out validation scene (or justify a scene-independent default) and report the main results under that protocol.
  3. [Eliminating Orientation Ambiguity, Eq. (5); Supplement §AxisNet] At inference the canonical rotation in Eq. (5) is driven by AxisNet's yaw estimate, yet no yaw-error statistics are reported anywhere in the paper or supplement. This matters because the w/o Align ablation in Table 5 (PCC 0.683 vs 0.738 full) shows orientation alignment carries a substantial share of the Layouts improvement, so the pipeline's sensitivity to yaw error is load-bearing for the CFAR claim. Note also an apparent mismatch: AxisNet is trained with ±90° yaw perturbations (supplement), while PFAL training uses only ±5° perturbations (Implementation Details), so the FloAff predictor may not be robust to realistic AxisNet errors. Please report AxisNet yaw MAE on the test scenes and a sensitivity curve of PCC/SIM/SR under injected yaw noise.
minor comments (8)
  1. [Benchmark Statistics; Table 8] Each task uses five training scenes and one unseen testing scene; test-set size is therefore small and the ± spreads in Tables 2–3 are presumably over rollout seeds rather than scenes. Please state what the ± quantities are computed over, and consider reporting per-scene variability or multiple test scenes.
  2. [Eq. (8)] Eq. (8) uses squared terms in the denominator (² + A²); this is the squared-soft-Dice variant, not the original V-Net form of Milletari et al. 2016 as cited. Either adjust the citation/wording or note the variant.
  3. [Table 2] Table 2: NavAff's Styles side-view SR (0.46) substantially exceeds its front-view SR (0.28), which is counterintuitive given that side views are described as harder. A brief discussion of this anomaly would help the reader.
  4. [Table 2] The Oracle row is defined only as 'MoMa performance at the optimal pose.' Please state precisely how the Oracle pose is chosen (e.g., argmax over the GT rollout-labeled map) and whether it uses information unavailable to the predictors.
  5. [Progressive FloAff Learning] The latent decomposition z = Φ(z_s, z_t) is introduced but never operationalized or measured; PFAL approximates it only 'conceptually.' Consider presenting it strictly as motivation, or provide a probe (e.g., feature-similarity analysis before/after adaptation with and without FD).
  6. [Supplement, Tables 6 and 8] Table 8 is captioned 'Statistics of RoboCasa dataset' but actually reports FloAff-Kitchen statistics; please retitle. Table 6 has inconsistent bold formatting across rows.
  7. [Progressive FloAff Priors Learning] The conceptual decomposition and progressive curriculum θ_pnp → θ_door,drawer suggests an ordering choice; an ablation on the adaptation order (e.g., starting from an articulated task, or adapting tasks one at a time vs jointly) would strengthen the PFAL claim, since the current evidence shows FD helps but does not isolate the curriculum itself.
  8. [Figure 4] Figure 4 is difficult to parse at print size; please enlarge the affordance-map insets and add a shared colorbar with numeric scale.

Circularity Check

1 steps flagged · score 3.0 of 10

Method training is ordinary supervised learning; mild circularity only in treating same-policy SR as independent proof beyond map fidelity.

  1. self definitional [Preliminaries Eq. (2); Supplement Eq. (16); Experiments Table 2 + Metrics]
    "the FloAff label is defined as: y_i = 1[Success(π_τ, b_i)], where y_i = 1 indicates successful task execution from pose b_i... we additionally report affordance-grounded manipulation Success Rate (SR) to evaluate the downstream MoMa performance enabled by FloAff prediction. ... Ours 0.85... 0.79... Oracle 0.88... 0.86"

    GT FloAff is literally the success map of the fixed task policies π_τ. Downstream SR places the robot using the predicted map and re-executes those same π_τ. Any method whose map closely matches GT therefore approaches Oracle SR by construction of the label—Table 2 largely re-expresses Table 1 map fidelity under the labeling policy, rather than independently verifying policy-agnostic physical affordance. The abstract claim that FloAff 'maximizes downstream manipulation success' is true of the labels by definition for this π family.

full rationale

CFAR (target crop + yaw alignment) and PFAL (PnP pretrain then staged adaptation with feature distillation) are standard representation and curriculum choices; nothing in Eqs. (3)–(11) forces the reported RMSE/PCC by algebraic identity with the inputs. Labels are rollout-defined (Eq. 2/16), the model is trained by MSE+SoftDice against those maps, and held-out scene metrics vs external baselines (NavAff, N2M, C2F-Exp, point-cloud backbones) are genuine empirical comparisons. The only load-bearing near-tautology is interpretive: FloAff is defined as Success(π_τ, b_i) for the same pretrained task policies later used to score FloAff-grounded MoMa SR, so a predictor that reproduces the GT map well is near-Oracle in Table 2 essentially because high-affordance cells are exactly the poses where π_τ already succeeded. That couples Table 2 to Table 1 and weakens the claim that SR is independent evidence of physical affordance beyond policy-specific standing regions—but it does not make the architecture or the cross-method ranking circular by construction. Crop radii chosen as the best rows of the parametric tables are ordinary hyperparameter selection, not a fitted constant renamed as a first-principles prediction. No uniqueness theorem or self-citation chain carries the central result. Score 3 reflects this single mild evaluation self-definition, not a forced derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The central empirical claim rests on domain modeling choices (local geometry sufficiency, PnP-as-foundation, policy-rollout labels), several hand-set hyperparameters (especially crop radius), and named modules (CFAR, PFAL, AxisNet) rather than on parameter-free theory. No new physical entities are postulated; the ledger is mostly robotics learning assumptions and fit knobs.

free parameters (4)
  • crop radius r = 1.30 m (Styles), 1.20 m (Layouts)
    Controls how much local geometry enters CFAR; selected via parametric sweeps and set differently for Styles (1.30 m) vs Layouts (1.20 m).
  • loss weights β and λ = β=1.0, λ=1.0
    Balance MSE vs Soft Dice and affordance loss vs feature-distillation term in PFAL; fixed by authors rather than derived.
  • PFAL/AxisNet optimization hyperparameters = PFAL: Adam 2e-4, 30 epochs, bs=8, yaw ±5°; AxisNet: Adam 1e-4, 200 epochs, bs=32, yaw ±90°, jitter σ=0.005
    Learning rates, epochs, batch sizes, and augmentation ranges are chosen for training stability and affect reported checkpoints.
  • candidate-pose search ranges and collision/reachability thresholds = task-dependent (not fully numeric in main text)
    Define which base poses are labeled at all; task-dependent rx, ry, r_arm and collision body approximations shape the GT maps.
assumptions (5)
  • domain assumption FloAff is primarily determined by target-conditioned local interaction geometry rather than the full observed scene.
    Stated in Introduction/CFAR and used to justify spherical cropping (Eq. 4); if distant structure or global layout dominates, CFAR discards needed signal.
  • domain assumption Yaw-only alignment to a canonical target facing direction makes affordance-equivalent scenes share a consistent representation for base placement.
    Eq. 5 rotates about the vertical axis using θ_o from sim (train) or AxisNet (test); assumes pitch/roll and non-yaw asymmetries are secondary.
  • domain assumption Binary success of a pretrained task policy at a base pose is a valid FloAff supervision signal after Gaussian smoothing.
    Eq. 2 and benchmark annotation section; couples labels to policy competence, horizon H, and simulator physics.
  • ad hoc to paper Hypothesis 1: heterogeneous MoMa skills share transferable base-placement priors that can be learned first from pick-and-place then adapted to hinge/sliding skills with feature distillation.
    Explicit hypothesis motivating PFAL stage order θ_pnp → door/drawer tasks and L_shared (Eqs. 6–11).
  • standard math Standard supervised learning losses (MSE + Soft Dice) on dense floor maps are appropriate objectives for FloAff prediction quality.
    Eqs. 7–8; conventional regression/overlap losses imported from dense prediction.
invented entities (3)
  • Canonical Floor Affordance Representation (CFAR)
    purpose: Named transform that crops target-centered points and aligns yaw to remove context/orientation ambiguity before FloAff prediction.
    Methodological construct defined in Definition 1 and Eqs. 3–5; not an external physical object, and evidence is only via ablations inside this paper’s benchmarks.
  • Progressive Floor Affordance Learning (PFAL)
    purpose: Curriculum that learns shared priors on PnP then adapts to articulated skills while distilling previous-stage features.
    Training procedure tied to Hypothesis 1; independent support would require external replications beyond reported tables.
  • FloAff-Kitchen benchmark (Styles/Layouts, multi-view)
    purpose: Provide cross-scene, front/side-view evaluation of FloAff and grounded MoMa in RoboCasa kitchens.
    New dataset resource claimed as first of its kind in this framing; value depends on public release and community uptake.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FloAff-Kitchen: Bridging Navigation and Manipulation via Canonical and Progressive Floor Affordance Learning." pith.science (2026). https://pith.science/paper/KO3Q5RGM

@misc{pith2026260724207,
  author       = {Pith},
  title        = {Pith review of: FloAff-Kitchen: Bridging Navigation and Manipulation via Canonical and Progressive Floor Affordance Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KO3Q5RGM}},
  note         = {Machine review of arXiv:2607.24207}
}
read the original abstract

Mobile manipulation requires robots to identify Floor Affordance (FloAff) that maximizes downstream manipulation success rather than merely ensuring navigation feasibility. FloAff prediction is a target-conditioned local spatial reasoning problem, yet existing methods suffer from representation ambiguity caused by irrelevant spatial context and arbitrary object orientations, while entangling shared and task-specific knowledge across heterogeneous manipulation skills. To address these challenges, we propose a unified framework for FloAff prediction from egocentric multimodal perception, consisting of canonical representation learning and progressive affordance prior learning. Specifically, we introduce a Canonical Floor Affordance Representation (CFAR), which learns canonical interaction geometry by preserving affordance-relevant local structure while eliminating nuisance spatial variations unrelated to robot base placement. We further propose Progressive Floor Affordance Learning (PFAL), which learns transferable FloAff priors from a foundation manipulation task and progressively adapts them to heterogeneous downstream manipulation skills. To facilitate systematic evaluation, we establish the first cross-scene, multi-view FloAff-Kitchen benchmark covering diverse manipulation skills, scene layouts, furniture styles, and viewpoints. Extensive experiments on three benchmark settings demonstrate that our method consistently outperforms strong baselines, while ablation studies validate the contribution of each proposed component. Project page: https://csu-hero-lab.github.io/FloAff-Kitchen_Web/

Figures

Figures reproduced from arXiv: 2607.24207 by the authors.

Figure 1
Figure 1. (Top) Examples of side-view and front-view MoMa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The data collection process consists of three phases: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustrations of (a) Canonical FloAff Representation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of qualitative results for predicting FloAff using different methods on the CloseD and OpenSD tasks [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Illustration of qualitative results for predicting FloAff using different methods on the PnPC2C and CloseDD tasks [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Illustration of FloAff predictions and the four down [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 7 linked inside Pith

  1. [1]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Moma-kitchen: A 100k+ benchmark for affordance-grounded last-mile navigation in mobile manipulation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  2. [2]

    Advances in neural information processing systems , volume=

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space , author=. Advances in neural information processing systems , volume=

  3. [3]

    proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Deep hough voting for 3d object detection in point clouds , author=. proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  4. [4]

    European conference on computer vision , pages=

    H3dnet: 3d object detection using hybrid geometric primitives , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  5. [5]

    arXiv preprint arXiv:2509.18671 , year=

    N2M: Bridging Navigation and Manipulation by Learning Pose Preference from Rollout , author=. arXiv preprint arXiv:2509.18671 , year=

    arXiv

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Affordance-guided coarse-to-fine exploration for base placement in open-vocabulary mobile manipulation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  7. [7]

    Robotics: Science and Systems (RSS) , year=

    RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots , author=. Robotics: Science and Systems (RSS) , year=

  8. [8]

    IEEE Transactions on Industrial Electronics , volume=

    ST-FMT*: A fast optimal global motion planning for mobile robot , author=. IEEE Transactions on Industrial Electronics , volume=. 2021 , publisher=

    2021

Show all 35 references
  1. [9]

    Science Robotics , volume=

    Navigating to objects in the real world , author=. Science Robotics , volume=. 2023 , publisher=

    2023

  2. [10]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    Think holistically, act down-to-earth: A semantic navigation strategy with continuous environmental representation and multi-step forward planning , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2023 , publisher=

    2023

  3. [11]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Treasure Hunting: Embodied Contrastive Learning-Enhanced Coarse-to-Fine Object Seeking With Explorer and Discriminator Cooperation , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

  4. [12]

    2020 IEEE international conference on robotics and automation (ICRA) , pages=

    Kimera: an open-source library for real-time metric-semantic localization and mapping , author=. 2020 IEEE international conference on robotics and automation (ICRA) , pages=. 2020 , organization=

    2020

  5. [13]

    arXiv preprint arXiv:2201.13360 , year=

    Hydra: A real-time spatial perception system for 3D scene graph construction and optimization , author=. arXiv preprint arXiv:2201.13360 , year=

    arXiv

  6. [14]

    IEEE Robotics and Automation Letters , volume=

    S-graphs+: Real-time localization and mapping leveraging hierarchical representations , author=. IEEE Robotics and Automation Letters , volume=. 2023 , publisher=

    2023

  7. [15]

    IEEE Robotics and Automation Letters , volume=

    Socially aware object goal navigation with heterogeneous scene representation learning , author=. IEEE Robotics and Automation Letters , volume=. 2024 , publisher=

    2024

  8. [16]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Hspnav: Hierarchical scene prior learning for visual semantic navigation towards real settings , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  9. [17]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Expand your scope: Semantic cognition over potential-based exploration for embodied visual navigation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  10. [18]

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

    Unigoal: Towards universal zero-shot goal-oriented navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [19]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    Chatnav: Leveraging llm to zero-shot semantic reasoning in object navigation , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2024 , publisher=

    2024

  12. [20]

    arXiv preprint arXiv:2601.14649 , year=

    Spatially Generalizable Mobile Manipulation via Adaptive Experience Selection and Dynamic Imagination , author=. arXiv preprint arXiv:2601.14649 , year=

  13. [21]

    2020 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Iris: Implicit reinforcement without interaction at scale for learning control from offline robot manipulation data , author=. 2020 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2020 , organization=

    2020

  14. [22]

    arXiv preprint arXiv:2304.13705 , year=

    Learning fine-grained bimanual manipulation with low-cost hardware , author=. arXiv preprint arXiv:2304.13705 , year=

    arXiv

  15. [23]

    The International Journal of Robotics Research , volume=

    Diffusion policy: Visuomotor policy learning via action diffusion , author=. The International Journal of Robotics Research , volume=. 2025 , publisher=

    2025

  16. [24]

    International Conference on Learning Representations , volume=

    Data scaling laws in imitation learning for robotic manipulation , author=. International Conference on Learning Representations , volume=

  17. [25]

    arXiv preprint arXiv:2601.16163 , year=

    Cosmos policy: Fine-tuning video models for visuomotor control and planning , author=. arXiv preprint arXiv:2601.16163 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2512.09297 , year=

    One-Shot Real-World Demonstration Synthesis for Scalable Bimanual Manipulation , author=. arXiv preprint arXiv:2512.09297 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2510.21571 , year=

    Scalable vision-language-action model pretraining for robotic manipulation with real-life human activity videos , author=. arXiv preprint arXiv:2510.21571 , year=

  20. [28]

    arXiv preprint arXiv:2507.17520 , year=

    Instructvla: Vision-language-action instruction tuning from understanding to manipulation , author=. arXiv preprint arXiv:2507.17520 , year=

  21. [29]

    IEEE Robotics and Automation Letters , volume=

    Robot learning of mobile manipulation with reachability behavior priors , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  22. [30]

    arXiv preprint arXiv:2209.02778 , year=

    Multi-skill mobile manipulation for object rearrangement , author=. arXiv preprint arXiv:2209.02778 , year=

    arXiv

  23. [31]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Navi2gaze: Leveraging foundation models for navigation and target gazing , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  24. [32]

    2013 IEEE International Conference on Robotics and Automation , pages=

    Robot placement based on reachability inversion , author=. 2013 IEEE International Conference on Robotics and Automation , pages=. 2013 , organization=

    2013

  25. [33]

    International Conference on Algorithms and Architectures for Parallel Processing , pages=

    MoMa-Pos: An Efficient Object-Kinematic-Aware Base Placement Determination Framework for Mobile Manipulation , author=. International Conference on Algorithms and Architectures for Parallel Processing , pages=. 2025 , organization=

    2025

  26. [34]

    2016 fourth international conference on 3D vision (3DV) , pages=

    V-net: Fully convolutional neural networks for volumetric medical image segmentation , author=. 2016 fourth international conference on 3D vision (3DV) , pages=. 2016 , organization=

    2016

  27. [35]

    arXiv preprint arXiv:1412.6980 , year=

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    arXiv

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.