RobotMover: Learning to Move Large Objects From Human Demonstrations

Akshara Rai; Alexander Clegg; Jimmy Yang; Joanne Truong; Sehoon Ha; Tianyu Li; Xavier Puig

REVIEW 4 major objections 7 minor 1 cited by

RobotMover: Learning to Move Large Objects From Human Demonstrations

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read RobotMover claims that a robot can learn to move large furniture by imitating a compact keypoint chain extracted from human-object demonstrations, with policies trained in simulation transferring directly to hardware.

desk verdict RobotMover is a clean chain-based imitation method with real Spot results; the hardware numbers are thin and the no-shape-randomization argument is shakier than the demos. read the letter →

arxiv 2502.05271 v2 pith:CELDP46P submitted 2025-02-07 cs.RO

Tianyu Li , Joanne Truong , Jimmy Yang , Alexander Clegg , Akshara Rai , Sehoon Ha , Xavier Puig This is my paper · ORCID

classification cs.RO

keywords largeobjectmanipulationmobileimitationlearninginteractionchaincross-embodimenttransferdomainrandomizationsim-to-realleggedrobot

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

RobotMover is an attempt to make large-object manipulation learnable without expensive robot teleoperation data or manually defined human-to-robot body maps. The paper's central claim is that a low-dimensional chain of keypoints, running from the human core through the active arm to the object, captures enough of the interaction for a robot to imitate: a policy rewarded for reproducing the same chain's motion learns to move the object. Training happens in simulation with randomized friction, mass, and starting poses, and the resulting policy is deployed on a quadruped robot with an onboard arm without fine-tuning. The paper reports that this policy outperforms four learning baselines in simulation and two learned plus two teleoperation baselines on hardware across capability, robustness, and controllability metrics, and that a chair-trained policy can even drag a loaded industrial trash cart. If these results hold, the approach points to a scalable route for teaching robots to handle furniture and other bulky objects in human environments.

What carries the argument

The central object is the Interaction Chain: a chain of keypoints anchored at the object's root and the agent's root, with intermediate nodes at the shoulder, elbow, and the agent-object contact point. The chain's state is the object's global position, orientation, and velocity plus the orientation of each chain segment in the object-plane frame, so it records both absolute motion and how force is transmitted from the body's core to the object. The same chain is extracted from human demonstrations and from the robot's actual contact, and the imitation reward is the negative exponential distance between the two chains, activated only while the robot's end-effector exerts contact force.

What would settle it

A concrete way to check the central claim: train a chair policy with the paper's single primitive shape and randomize friction and mass exactly as described, then deploy it zero-shot on a real chair with the same mass and friction but a visibly different contact geometry, such as a much thinner backrest. If the chair is dropped or collides with the robot much more often than the reported success rates, the premise that geometry reduces to dynamic randomization is wrong.

Watch

Extended reading notes

Core claim

The paper's discovery, stated on its own terms, is that using the Interaction Chain as the imitation reference eliminates the need for manually defining whole-body correspondences between the human and the robot, enabling more scalable and transferable policy learning. The Interaction Chain is a chain of keypoints connecting the agent's root to the object's root through the shoulder, elbow, and contact point, and it captures both the object's motion and how force is transmitted from the agent's core to the object. The same chain is extracted from human demonstrations and from the robot's dynamically formed contact, and the imitation reward penalizes the distance between them only while the robot maintains contact force with the object. The paper's experimental conclusion is that RobotMover outperforms all baselines over all metrics on hardware, including maximum stable velocity, robustness to object diversity and initial conditions, and trajectory tracking accuracy.

Load-bearing premise

The load-bearing premise is that a policy trained on one simplified object shape per category generalizes to real objects of that category because shape differences only show up as dynamic differences after grasp, which are covered by randomizing friction, mass, and starting pose.

Editorial extensions

If this is right

A simulation-trained policy for chairs, tables, and standing racks transfers zero-shot to a real quadruped robot with an onboard arm, with no fine-tuning after deployment.
The learned policy tracks commanded object velocities well enough to be steered by a high-level planner or a human operator for long-horizon transport and rearrangement tasks.
The interaction chain reward removes the need to manually define whole-body human-to-robot correspondence, which the paper identifies as a bottleneck for Interaction Graph-based imitation.
A chair-trained policy adapted to a much larger loaded trash cart without retraining, suggesting that the learned dynamics are not tightly tied to one object's exact shape.
Two-arm human demonstrations can support policy learning either by aggregating the two chains into one reference or by splitting the chains across two collaborating robots.

Reading between the lines

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

Editorial inference: because the policy's input is a commanded object velocity rather than a visual or object-state observation, any high-level planner that outputs object velocity waypoints could drive the same trained policy, enabling whole-room rearrangement without new perception hardware.
Editorial inference: the Interaction Chain's node count is naturally flexible, so the representation could extend beyond arm-driven manipulation to pushing, towing, or torso-contact tasks where the chain would run through different body parts.
Editorial inference: the paper trains a separate policy per object category; a direct extension would be to train a single policy across categories and test whether the chain representation alone can absorb cross-object dynamics, which the current experiments do not yet establish.

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.

Desk Editor's Note

read the letter

Here's my take on RobotMover. The useful thing in this paper is the Interaction Chain: a deliberately stripped-down version of Interaction Graphs that connects object root to agent root through a handful of keypoints. Using that as an imitation reward is a clean idea, and the paper shows it works on a real Spot moving chairs, tables, and racks, with zero-shot sim-to-real transfer. The hardware demos, especially the trash cart and the chair rearrangement, are genuinely useful evidence that the method works outside the training distribution.

The paper is also honest about its limits. It discloses the manual grasp initialization, the lack of visual feedback, and the per-category policy training. That is refreshing and makes the claims easier to evaluate.

The main soft spots are statistical and interpretive. The hardware numbers in Figure 8 are point values with small trial counts (5–10 per condition), no error bars, no significance tests. For a claim like 'outperforms all baselines over all metrics,' that is thin. The controllability metrics (STT-CON, MTT-CON) are underspecified in the text; I could not tell exactly how the endpoint error was computed.

The bigger interpretive issue is the Sec V-B claim that shape randomization is unnecessary because geometric differences are ultimately reflected in dynamic variations. That physical argument does not quite hold: contact geometry affects where forces can act, and mass/friction randomization will not span the difference between a thin tubular back and a wide flat one. The paper's own evidence is mixed—the thin chair and trash cart work, which is good, but that success is not explained by the friction-randomization story. It suggests the policy is robust to dynamics perturbations in ways the paper does not fully account for. This does not sink the paper, but it means the generalization claim should be read more narrowly than the abstract suggests.

The circularity concern is real but mild: the imitation reward includes object trajectory distance, which is essentially the evaluation metric. The baselines share that term, so the comparison is still informative, but part of the reported success is aligned with the training objective.

If I were editing, I would send this to review. The core idea is new enough, the hardware results are meaningful despite the small trial counts, and the limitations are disclosed. The authors should be asked to add error bars, clarify the metrics, and either soften or better support the shape-randomization claim. Worth reading if you work in mobile manipulation or sim-to-real.

Referee Report

4 major / 7 minor

Summary. RobotMover proposes an imitation-learning framework for whole-body manipulation of large objects by a quadrupedal mobile manipulator. Human-object interaction demonstrations from the OMOMO dataset are mapped to a low-dimensional “Interaction Chain” representation connecting the object root to the agent root through contact and body keypoints. The distance between the human and robot chains in Eq. (3) defines an imitation reward, and policies are trained with PPO in two Genesis simulation settings under domain randomization, then deployed zero-shot on a Boston Dynamics Spot robot. The paper reports that the learned policy tracks object velocities better than RL, RL-EE, RL-IK, and RL-IG baselines in simulation, and that it outperforms learning and teleoperation baselines on six hardware metrics across chairs, tables, and racks. Two long-horizon applications are demonstrated: trash-cart transport and chair rearrangement.

Significance. If the results hold, the paper would make a useful practical contribution: a morphology-agnostic imitation representation that avoids hand-specified body correspondences, a sim-to-real pipeline for dynamic manipulation of bulky objects on a legged robot, and a substantial real-world evaluation. The use of a public motion-capture dataset, the explicit domain-randomization protocol, and the hardware experiments with multiple object categories are strengths. The proposed Interaction Chain is a simple and plausible idea, and the application demonstrations are compelling. However, the headline claims are not yet fully established because the hardware evaluation lacks statistical support and because the zero-shot generalization claim rests on an unverified physical premise about shape variation. The contribution is potentially significant but needs revision to match the evidence.

major comments (4)

[VI-D] The claim that “RobotMover outperforms all baselines over all metrics” is supported only by point estimates. The protocols in Section VI-C use 5 runs for the velocity-capability metrics, 10 trials per object for DIV-ROB, and 30 trials for IC-ROB, and none of the hardware results include confidence intervals, variances, or significance tests. With these sample sizes, the reported differences may be within noise, particularly for the controllability metrics. Please report per-condition trial counts, error bars, and a statistical comparison (for example, bootstrap confidence intervals or paired tests), or explicitly soften the claim to a qualitative demonstration.
[V-B / VI-E] The justification for not randomizing object shape is load-bearing for the zero-shot generalization claim. Section V-B states that “geometric differences are ultimately reflected in dynamic variations once the object is grasped, and these dynamics can be effectively randomized through variations in friction, mass, and initial poses.” This premise is not established. Contact geometry determines the contact normal, contact area, pressure distribution, and lever arms about the gripper, so two objects with identical mass, friction, and initial pose can behave very differently. The thin-chair anecdote in Section VI-E is a single counterexample, and the trash-cart application uses a wheeled object whose rolling and caster dynamics are not represented by the sliding-friction randomization of the simulated primitive chair. Please add systematic shape-generalization experiments, or explicitly restrict the generalization claim to dynamic variation and present the trash-cart and thin-chair results as demonstrations of robustness rather than evidence for the no-shape-randomization premise.
[IV-E / VI-A] The simulation evaluation metric in Eq. (5), vtrack = (1/T) Σ exp(-|x_ho,xy - x_ro,xy|), is a monotone transform of the object-trajectory term ‖x_ho - x_ro‖ that appears directly in the RobotMover imitation reward in Eq. (3). Thus the primary simulation metric is aligned with the training objective by construction. This does not invalidate the comparison because the baselines also receive object-trajectory guidance, but it means the simulation results are partially circular with respect to the claimed advantage of the Interaction Chain. Please report at least one evaluation metric that is not contained in the reward (for example, success rate under perturbations, contact-force smoothness, or planned-trajectory endpoint error) and discuss the overlap explicitly.
[VI-C / VII] The hardware evaluation begins after the robot is manually teleoperated to a stable grasp, and the paper acknowledges in Section VII that grasping is outside the current framework. This is an honest limitation, but it conflicts with the abstract's description of “a complete learning-based system” and with the framing that the policy learns to “move” objects in an end-to-end sense. Please make this scope explicit in the abstract and contributions, or add experiments that include autonomous approach and grasp acquisition. The current scope is still publishable as a moving-phase controller, but the claims should match that scope.

minor comments (7)

[Fig. 3] The caption in Figure 3 refers to the “Dynamic Chain” instead of the “Interaction Chain”; please fix this inconsistent terminology.
[Eq. (3)] The definition of N is unclear: the text says “N denotes the number of nodes between the robot’s and the human’s Interaction Chains,” but the sum runs from i=0 to N-1, suggesting N segments or N-1 intermediate nodes. Please clarify the indexing and state explicitly whether α_0 is the weight for the object-root segment.
[VI-B] The statement that “RobotMover outperforms all baselines in almost every setting” conflicts with the observation that RL-EE is comparable to RobotMover for some chair and table experiments, and with the stronger wording used later in Section VI-D. Please report the per-condition numbers and reconcile the wording.
[VI-C] The definitions of DIV-ROB and of the trajectory-tracking metrics are hard to parse. The formula “vDIV-ROB = 1/(10N) Σ ... 1(if success)” is missing a limit or an explicit sum, and the text has typos such as “the metric measure also measure the motion transition property.” Please rewrite the metric definitions with clear notation and complete sentences.
[Appendix C] Algorithm 1 has a bracketing typo on line 10 (min(0.4, δhead] has mismatched parentheses), and the sign convention for “the delta heading towards the target position” is not defined.
[Fig. 12] The object-size table in Figure 12 lists sizes without units in the header; the units appear to be centimeters, but this should be stated explicitly.
[References] Reference [25] is cited as “Ravan et al.”; please verify the author spelling and ensure the citation matches the published record.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild reward/evaluation overlap in the object-tracking term; the central Interaction Chain comparison is independently supported.

other [Section IV-E, Eq. (3) and Section VI-A, Eq. (5)]
"Eq. (3): `errc_t = ||xho_t − xro_t || + Σ α_i∥qho,i_t − qro,i_t∥`, where `|xho_t − xro_t | measures the difference between the object movements induced by the human and the robot`. Eq. (5): `vtrack = (1/T) Σ e^{−|xho,xy_t − xro,xy_t |}`, where `xho,xy_t and xro,xy_t denote the 2D xy-positions and headings of the reference and experimental objects at time t`."

The first, highest-weighted term of the imitation reward is the distance between the human-demonstration object pose and the robot object pose; the simulation evaluation metric is the exponential of the same object-pose distance (projected to xy and heading). Therefore the absolute vtrack numbers reported in Fig. 7 are not an independent test of object-tracking: the policy is trained to maximize essentially the same quantity that is later measured. This is a partial alignment by construction. It does not undermine the central comparison because every learning baseline (RL, RL-EE, RL-IK, RL-IG) is also rewarded/trained against the reference object trajectory, so the incremental advantage of RobotMover is still attributable to the added Interaction Chain terms.

full rationale

RobotMover's central derivation is not circular. The Interaction Chain is a new spatial descriptor defined in Section IV-B from human and robot keypoints; the imitation reward in Eq. (3) compares chain node positions and segment orientations; and the simulation experiments in Section VI-B compare this reward against four baselines under the same object-trajectory supervision, so the reported advantage isolates the chain representation rather than being forced by the evaluation metric. The only overlap pinned to the paper's own equations is that the reward's object-position term is the same quantity exponentiated in the tracking metric vtrack (Eq. 5), which means the absolute tracking scores are partly aligned with training; this is mild and does not affect the relative comparison. Self-citations ([27], [29], [30], [33]) concern prior motion retargeting and locomotion methods and are not used as a load-bearing justification for the Interaction Chain or for excluding alternatives. The Section V-B assertion that shape randomization is unnecessary is a physical assumption with only anecdotal support; it is a generalization-risk concern, not a circularity.

Assumptions & free parameters 4 free parameters · 7 assumptions · 1 invented entities

The central claim rests on the fidelity of simulation to real hardware, the sufficiency of the OMOMO human demonstrations, and the assumption that mass, friction, and pose randomization can substitute for shape variation. None of these is independently verified outside the paper. The Interaction Chain itself is an invented representational tool with no external falsifiable handle.

free parameters (4)

alpha_i (chain segment weights)
Hand-assigned in Eq. (3), with higher weights near the object. Values are not reported in the paper, and the reward balance between object tracking and arm configuration depends on them.
Imitation and regularization reward coefficients
The paper mentions height, torque, rotation, and action-rate penalties following [40], but does not report the coefficient values. These determine motion quality and policy behavior.
Domain randomization ranges
The ranges for object mass, friction, initial pose, and robot height are not specified. The zero-shot transfer claim hinges on these ranges being wide enough to cover real-world variation.
PPO hyperparameters
Learning rate, discount factor, GAE lambda, clip ratio, and other PPO settings are not reported. These are standard free choices that affect convergence and final performance.

assumptions (7)

standard math Reinforcement learning with PPO converges to a policy that maximizes the imitation reward in simulation.
Standard RL assumption; the paper uses PPO [41] and reports convergence in about 4 hours.
domain assumption OMOMO motion capture demonstrations provide suitable references for human-object interaction dynamics for the target objects.
Section V-A and Appendix IX-A: demonstrations are taken from OMOMO. No validation is given that the human strategies are sufficient or optimal for the robot to imitate.
domain assumption Primitive-geometry object models with randomized mass, friction, and initial pose capture the real-world dynamics of the target objects, so shape randomization is unnecessary.
Section V-B states this explicitly. It is load-bearing for generalization to the thin chair and trash cart, and it is challenged by the fact that only one geometric model per object type is used.
domain assumption The simplified dynamics environment, where the policy commands root velocity and arm pose targets, is a valid proxy for the real Spot robot's closed-loop locomotion.
Section V-B: the deployed policy uses only the simplified dynamics interface. Failures of Spot's internal controller under load would break the transfer claim.
ad hoc to paper The Interaction Chain, built from the dominant contact points, captures the essential dynamics for imitation; nodes and edges not in the chain are irrelevant.
Section IV-B: the chain is designed to be compact. The paper provides no independent criterion for choosing which keypoints to include and which to omit.
domain assumption A reward that only activates when contact force is positive (Eq. 4) is sufficient to shape the desired manipulation behavior.
Section IV-E: the contact-gated reward assumes contact can be reliably sensed and is always desirable. No sensitivity analysis is provided.
domain assumption The policy can control object motion without observing object pose or velocity, relying only on proprioception, gripper contact, and target velocity.
Section V-C: the observation space excludes object state and visual input. This is load-bearing for the generalization claim, since the policy must infer task state from limited signals.

invented entities (1)

Interaction Chain
purpose: A morphology-agnostic spatial descriptor, a chain of keypoints from agent root to object root, used to define the imitation reward and to compare human and robot interactions.
This is a representational construct, not a physical entity. Its utility is established only through the paper's experiments; no external falsifiable prediction is made.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RobotMover: Learning to Move Large Objects From Human Demonstrations." pith.science (2026). https://pith.science/paper/CELDP46P

@misc{pith2026250205271,
  author       = {Pith},
  title        = {Pith review of: RobotMover: Learning to Move Large Objects From Human Demonstrations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CELDP46P}},
  note         = {Machine review of arXiv:2502.05271}
}

read the original abstract

Moving large objects, such as furniture or appliances, is a critical capability for robots operating in human environments. This task presents unique challenges, including whole-body coordination to avoid collisions and managing the dynamics of bulky, heavy objects. In this work, we present RobotMover, a learning-based system for large object manipulation that uses human-object interaction demonstrations to train robot control policies. RobotMover formulates the manipulation problem as imitation learning using a simplified spatial representation called the Interaction Chain, which captures essential interaction dynamics in a way that generalizes across different robot bodies. We incorporate this Interaction Chain into a reward function and train policies in simulation using domain randomization to enable zero-shot transfer to real-world robots. The resulting policies allow a Spot robot to manipulate various large objects, including chairs, tables, and standing lamps. Through extensive experiments in both simulation and the real world, we show that RobotMover achieves strong performance in terms of capability, robustness, and controllability, outperforming both learned and teleoperation baselines. The system also supports practical applications by combining learned policies with simple planning modules to perform long-horizon object transport and rearrangement tasks.

Figures

Figures reproduced from arXiv: 2502.05271 by the authors.

**Figure 2.** Challenges of moving large objects include but not [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Method Overview. RobotMover enables robots to learn [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Interaction Graph (top) vs. Interaction Chain (bottom). [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 6.** Figure 6: Two simulation settings. Left: simple dynamic envi [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Quantitative results of simulation experiments. Here, ‘SD’ represents simple-dynamics environment while ‘FD’ indicates [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Summary of Hardware Quantitative Results. We compare our proposed [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 9.** Figure 9: Robot moving a chair using different methods. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

**Figure 10.** Figure 10: Illustration of two applications: Trash Cart Transportation (top) and Chair Rearrangement (bottom). A motion planner [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

**Figure 11.** Figure 11: Two robots learning to move a coffee table from two [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

**Figure 12.** Figure 12: The collection of objects used in the diverse object [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

SLAC: Safe and Efficient Real-Robot Reinforcement Learning via Unsupervised Simulation Pre-Training
cs.RO 2025-06 conditional novelty 5.0 of 10

SLAC learns a latent action space in a low-fidelity simulator and uses it for real-world reinforcement learning, solving whole-body mobile manipulation tasks in under an hour without demonstrations.

Reference graph

Works this paper leans on

44 extracted references · 35 canonical work pages · cited by 1 Pith paper

[1]

Do as i can, not as i say: Grounding language in robotic affordances,

M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Haus- man et al., “Do as i can, not as i say: Grounding language in robotic affordances,” Conference on Robot Learning , 2022

work page 2022
[2]

Diffusion pol- icy: Visuomotor policy learning via action diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion pol- icy: Visuomotor policy learning via action diffusion,” The International Journal of Robotics Research , p. 02783649241273668, 2023. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, XXXXXXX 202X 13

work page 2023
[3]

pi 0: A vision-language-action flow model for general robot control,

K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter et al. , “ pi 0: A vision-language-action flow model for general robot control,” arXiv preprint arXiv:2410.24164, 2024

arXiv 2024
[4]

Octo: An open-source generalist robot policy,

O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu et al., “Octo: An open-source generalist robot policy,” arXiv preprint arXiv:2405.12213, 2024

arXiv 2024
[5]

Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” Conference on Robot Learning (CoRL) , 2024

work page 2024
[6]

Mobile-television: Predictive mo- tion priors for humanoid whole-body control,

C. Lu, X. Cheng, J. Li, S. Yang, M. Ji, C. Yuan, G. Yang, S. Yi, and X. Wang, “Mobile-television: Predictive mo- tion priors for humanoid whole-body control,” IEEE International Conference on Robotics and Automation (ICRA), 2025

work page 2025
[7]

Tidybot: Personalized robot assistance with large lan- guage models,

J. Wu, R. Antonova, A. Kan, M. Lepert, A. Zeng, S. Song, J. Bohg, S. Rusinkiewicz, and T. Funkhouser, “Tidybot: Personalized robot assistance with large lan- guage models,” Autonomous Robots, vol. 47, no. 8, pp. 1087–1102, 2023

work page 2023
[8]

Asc: Adaptive skill coordination for robotic mobile manipulation,

N. Yokoyama, A. Clegg, J. Truong, E. Undersander, T.-Y . Yang, S. Arnaud, S. Ha, D. Batra, and A. Rai, “Asc: Adaptive skill coordination for robotic mobile manipulation,” IEEE Robotics and Automation Letters , vol. 9, no. 1, pp. 779–786, 2023

work page 2023

Show all 44 references

[9]

Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environ- ments,

S. Srivastava, C. Li, M. Lingelbach, R. Mart ´ın-Mart´ın, F. Xia, K. E. Vainio, Z. Lian, C. Gokmen, S. Buch, K. Liu et al. , “Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environ- ments,” in Conference on robot learning. PMLR, 2...

2022
[10]

Habitat: A platform for embodied ai research,

M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wi- jmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Malik et al., “Habitat: A platform for embodied ai research,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 9339–9347

2019
[11]

Articulated object interaction in unknown scenes with whole-body mobile manipulation,

M. Mittal, D. Hoeller, F. Farshidian, M. Hutter, and A. Garg, “Articulated object interaction in unknown scenes with whole-body mobile manipulation,” in 2022 IEEE/RSJ international conference on intelligent robots and systems (IROS) . IEEE, 2022, pp. 1647–1654

2022
[12]

Deep whole-body control: learning a unified policy for manipulation and locomotion,

Z. Fu, X. Cheng, and D. Pathak, “Deep whole-body control: learning a unified policy for manipulation and locomotion,” in Conference on Robot Learning. PMLR, 2023, pp. 138–149

2023
[13]

Rt-1: Robotics transformer for real-world control at scale,

A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu et al., “Rt-1: Robotics transformer for real-world control at scale,” arXiv preprint arXiv:2212.06817, 2022

2022 arXiv
[14]

The darpa robotics challenge finals: Results and perspectives,

E. Krotkov, D. Hackett, L. Jackel, M. Perschbacher, J. Pippine, J. Strauss, G. Pratt, and C. Orlowski, “The darpa robotics challenge finals: Results and perspectives,” The DARPA robotics challenge finals: Humanoid robots to the rescue , pp. 1–26, 2018

2018
[15]

Towards a personal robotics develop- ment platform: Rationale and design of an intrinsically safe personal robot,

K. A. Wyrobek, E. H. Berger, H. M. Van der Loos, and J. K. Salisbury, “Towards a personal robotics develop- ment platform: Rationale and design of an intrinsically safe personal robot,” in 2008 IEEE International Con- ference on Robotics and Automation . IEEE, 2008, pp. 2165–2170

2008
[16]

Force strategies for cooper- ative tasks in multiple mobile manipulation systems,

O. Khatib, K. Yokoi, K. Chang, D. Ruspini, R. Holmberg, A. Casal, and A. Baader, “Force strategies for cooper- ative tasks in multiple mobile manipulation systems,” in Robotics Research: The Seventh International Sym- posium. Springer, 1996, pp. 333–342

1996
[17]

Integrated task and motion planning,

C. R. Garrett, R. Chitnis, R. Holladay, B. Kim, T. Silver, L. P. Kaelbling, and T. Lozano-P ´erez, “Integrated task and motion planning,” Annual review of control, robotics, and autonomous systems , vol. 4, no. 1, pp. 265–293, 2021

2021
[18]

Fully autonomous real-world reinforcement learning with applications to mobile ma- nipulation,

C. Sun, J. Orbik, C. M. Devin, B. H. Yang, A. Gupta, G. Berseth, and S. Levine, “Fully autonomous real-world reinforcement learning with applications to mobile ma- nipulation,” in Conference on Robot Learning . PMLR, 2022, pp. 308–319

2022
[19]

Homerobot: Open-vocabulary mobile manipulation,

S. Yenamandra, A. Ramachandran, K. Yadav, A. Wang, M. Khanna, T. Gervet, T.-Y . Yang, V . Jain, A. W. Clegg, J. Turner et al. , “Homerobot: Open-vocabulary mobile manipulation,” arXiv preprint arXiv:2306.11565 , 2023

2023 arXiv
[20]

Umi on legs: Making manipulation policies mobile with manipulation- centric whole-body controllers,

H. Ha, Y . Gao, Z. Fu, J. Tan, and S. Song, “Umi on legs: Making manipulation policies mobile with manipulation- centric whole-body controllers,” Conference on Robot Learning (CoRL), 2024

2024
[21]

Adap- tive mobile manipulation for articulated objects in the open world,

H. Xiong, R. Mendonca, K. Shaw, and D. Pathak, “Adap- tive mobile manipulation for articulated objects in the open world,” arXiv preprint arXiv:2401.14403 , 2024

2024 arXiv
[22]

On bringing robots home,

N. M. M. Shafiullah, A. Rai, H. Etukuru, Y . Liu, I. Misra, S. Chintala, and L. Pinto, “On bringing robots home,” arXiv preprint arXiv:2311.16098 , 2023

2023 arXiv
[23]

Moma-force: Visual-force im- itation for real-world mobile manipulation,

T. Yang, Y . Jing, H. Wu, J. Xu, K. Sima, G. Chen, Q. Sima, and T. Kong, “Moma-force: Visual-force im- itation for real-world mobile manipulation,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 6847–6852

2023
[24]

Continuously improving mobile manipu- lation with autonomous real-world rl,

R. Mendonca, E. Panov, B. Bucher, J. Wang, and D. Pathak, “Continuously improving mobile manipu- lation with autonomous real-world rl,” Conference on Robot Learning (CoRL) , 2024

2024
[25]

Combining planning and diffusion for mo- bility with unknown dynamics,

Y . Ravan, Z. Yang, T. Chen, T. Lozano-P ´erez, and L. P. Kaelbling, “Combining planning and diffusion for mo- bility with unknown dynamics,” 2023 IEEE International Conference on Robotics and Automation (ICRA) , 2025

2023
[26]

Bc-z: Zero-shot task generalization with robotic imitation learning,

E. Jang, A. Irpan, M. Khansari, D. Kappler, F. Ebert, C. Lynch, S. Levine, and C. Finn, “Bc-z: Zero-shot task generalization with robotic imitation learning,” in Conference on Robot Learning . PMLR, 2022, pp. 991– 1002

2022
[27]

Crossloco: Human motion driven control of legged robots via guided unsupervised reinforcement learning,

T. Li, H. Jung, M. Gombolay, Y . K. Cho, and S. Ha, “Crossloco: Human motion driven control of legged robots via guided unsupervised reinforcement learning,” International Conference on Learning Representations , 2024

2024
[28]

Behavior transformers: Cloning k modes with one stone,

N. M. Shafiullah, Z. Cui, A. A. Altanzaya, and L. Pinto, JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, XXXXXXX 202X 14 “Behavior transformers: Cloning k modes with one stone,” Advances in neural information processing sys- tems, vol. 35, pp. 22 955–22 968, 2022

2022
[29]

Ace: Adversarial correspondence embedding for cross morphology motion retargeting from human to nonhu- man characters,

T. Li, J. Won, A. Clegg, J. Kim, A. Rai, and S. Ha, “Ace: Adversarial correspondence embedding for cross morphology motion retargeting from human to nonhu- man characters,” in SIGGRAPH Asia 2023 Conference Papers, 2023, pp. 1–11

2023
[30]

Learning and adapting agile locomotion skills by transferring experience,

L. Smith, J. C. Kew, T. Li, L. Luu, X. B. Peng, S. Ha, J. Tan, and S. Levine, “Learning and adapting agile locomotion skills by transferring experience,” Robotics: Science and Systems , 2023

2023
[31]

Open x- embodiment: Robotic learning datasets and rt-x models,

Q. Vuong, S. Levine, H. R. Walke, K. Pertsch, A. Singh, R. Doshi, C. Xu, J. Luo, L. Tan, D. Shah et al., “Open x- embodiment: Robotic learning datasets and rt-x models,” in Towards Generalist Robots: Learning Paradigms for Scalable Skill Acquisition@ CoRL2023 , 2023

2023
[32]

Universal manipu- lation interface: In-the-wild robot teaching without in- the-wild robots,

C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipu- lation interface: In-the-wild robot teaching without in- the-wild robots,” Robotics: Science and Systems , 2024

2024
[33]

Using deep reinforcement learning to learn high-level policies on the atrias biped,

T. Li, H. Geyer, C. G. Atkeson, and A. Rai, “Using deep reinforcement learning to learn high-level policies on the atrias biped,” in 2019 International Conference on Robotics and Automation (ICRA) . IEEE, 2019, pp. 263–269

2019
[34]

Learning agile robotic locomotion skills by imitating animals,

X. B. Peng, E. Coumans, T. Zhang, T.-W. E. Lee, J. Tan, and S. Levine, “Learning agile robotic locomotion skills by imitating animals,” in Robotics: Science and Systems , 07 2020

2020
[35]

Deepmimic: Example-guided deep reinforcement learn- ing of physics-based character skills,

X. B. Peng, P. Abbeel, S. Levine, and M. Van de Panne, “Deepmimic: Example-guided deep reinforcement learn- ing of physics-based character skills,” ACM Transactions On Graphics (TOG) , vol. 37, no. 4, pp. 1–14, 2018

2018
[36]

Amp: Adversarial motion priors for styl- ized physics-based character control,

X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “Amp: Adversarial motion priors for styl- ized physics-based character control,” ACM Transactions on Graphics (ToG), vol. 40, no. 4, pp. 1–20, 2021

2021
[37]

Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters,

X. B. Peng, Y . Guo, L. Halper, S. Levine, and S. Fidler, “Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters,” ACM Transactions On Graphics (TOG) , vol. 41, no. 4, pp. 1–17, 2022

2022
[38]

Simulation and retargeting of complex multi-character interactions,

Y . Zhang, D. Gopinath, Y . Ye, J. Hodgins, G. Turk, and J. Won, “Simulation and retargeting of complex multi-character interactions,” in ACM SIGGRAPH 2023 Conference Proceedings, 2023, pp. 1–11

2023
[39]

Spatial relationship preserving character motion adaptation,

E. S. Ho, T. Komura, and C.-L. Tai, “Spatial relationship preserving character motion adaptation,” in ACM SIG- GRAPH 2010 papers , 2010, pp. 1–8

2010
[40]

Learning to walk in minutes using massively parallel deep rein- forcement learning,

N. Rudin, D. Hoeller, P. Reist, and M. Hutter, “Learning to walk in minutes using massively parallel deep rein- forcement learning,” in Conference on Robot Learning . PMLR, 2022, pp. 91–100

2022
[41]

Proximal policy optimization algorithms,

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347 , 2017

2017 arXiv
[42]

Object motion guided hu- man motion synthesis,

J. Li, J. Wu, and C. K. Liu, “Object motion guided hu- man motion synthesis,” ACM Transactions on Graphics (TOG), vol. 42, no. 6, pp. 1–11, 2023

2023
[43]

Genesis: A universal and generative physics engine for robotics and beyond,

G. Authors, “Genesis: A universal and generative physics engine for robotics and beyond,” Decem- ber 2024. [Online]. Available: https://github.com/ Genesis-Embodied-AI/Genesis

2024
[44]

Fast and accurate deep network learn- ing by exponential linear units (elus),

D.-A. Clevert, “Fast and accurate deep network learn- ing by exponential linear units (elus),” arXiv preprint arXiv:1511.07289, 2015. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, XXXXXXX 202X 15 IX. A PPENDIX A. Dataset Details We selected human-object interaction demonstrati...

2015 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Do as i can, not as i say: Grounding language in robotic affordances,

M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Haus- man et al., “Do as i can, not as i say: Grounding language in robotic affordances,” Conference on Robot Learning , 2022

work page 2022

[2] [2]

Diffusion pol- icy: Visuomotor policy learning via action diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion pol- icy: Visuomotor policy learning via action diffusion,” The International Journal of Robotics Research , p. 02783649241273668, 2023. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, XXXXXXX 202X 13

work page 2023

[3] [3]

pi 0: A vision-language-action flow model for general robot control,

K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter et al. , “ pi 0: A vision-language-action flow model for general robot control,” arXiv preprint arXiv:2410.24164, 2024

arXiv 2024

[4] [4]

Octo: An open-source generalist robot policy,

O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu et al., “Octo: An open-source generalist robot policy,” arXiv preprint arXiv:2405.12213, 2024

arXiv 2024

[5] [5]

Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” Conference on Robot Learning (CoRL) , 2024

work page 2024

[6] [6]

Mobile-television: Predictive mo- tion priors for humanoid whole-body control,

C. Lu, X. Cheng, J. Li, S. Yang, M. Ji, C. Yuan, G. Yang, S. Yi, and X. Wang, “Mobile-television: Predictive mo- tion priors for humanoid whole-body control,” IEEE International Conference on Robotics and Automation (ICRA), 2025

work page 2025

[7] [7]

Tidybot: Personalized robot assistance with large lan- guage models,

J. Wu, R. Antonova, A. Kan, M. Lepert, A. Zeng, S. Song, J. Bohg, S. Rusinkiewicz, and T. Funkhouser, “Tidybot: Personalized robot assistance with large lan- guage models,” Autonomous Robots, vol. 47, no. 8, pp. 1087–1102, 2023

work page 2023

[8] [8]

Asc: Adaptive skill coordination for robotic mobile manipulation,

N. Yokoyama, A. Clegg, J. Truong, E. Undersander, T.-Y . Yang, S. Arnaud, S. Ha, D. Batra, and A. Rai, “Asc: Adaptive skill coordination for robotic mobile manipulation,” IEEE Robotics and Automation Letters , vol. 9, no. 1, pp. 779–786, 2023

work page 2023