Learning Dexterous Object Handover

Daniel Frau-Alfaro; Julio Casta\~no-Amoros; Pablo Gil; Roberto Calandra; Santiago Puente

REVIEW 4 major objections 6 minor 23 references

Learning Dexterous Object Handover

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A dual-quaternion reward lets a reinforcement-trained four-finger hand outperform an Euler-angle reward for robot-to-robot object handover in simulation, reaching a 94% best-case total success rate.

desk verdict The DQ-vs-Euler result is probably real, but the headline 91-94% success numbers are inflated by counting simulator collision failures as success. read the letter →

arxiv 2506.16822 v1 pith:XYBT2ANJ submitted 2025-06-20 cs.RO cs.AI

Daniel Frau-Alfaro , Julio Castaño-Amoros , Santiago Puente , Pablo Gil , Roberto Calandra This is my paper · ORCID

classification cs.ROcs.AI

keywords reinforcementlearningobjecthandoverdualquaternionsrewardfunctiondesigndexterousmanipulationmulti-fingerhandsrotationdistancerobot-robotcollaboration

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

This paper tries to show that the choice of rotation representation inside a reinforcement-learning reward function decides whether a four-finger robot hand can learn to take an object from another robot hand. Its central proposal is a reward term that measures pose distance with dual quaternions, which encode rotation and translation in one object, instead of Euler angles or rotation matrices. In simulation, the dual-quaternion agent reaches a 91% total success rate on the training object and 94% on a shorter unseen prism, beating the Euler-angle agent's 83% and 92%, while a rotation-matrix reward fails to learn at all. If correct, this would make dual quaternions a practical choice for the reward geometry of dexterous handover tasks, not just a mathematical curiosity.

What carries the argument

The load-bearing object is dual quaternion algebra for poses in $SE(3)$: a pose is written $\hat{q} = q_r + \varepsilon \tfrac{1}{2} q_r q_t$, with $q_r$ the rotation quaternion and $q_t$ the translation expressed as a pure quaternion. The distance used in the reward is $d_{\mathrm{DQ}} = \|\hat{q}_{\mathrm{diff}} - \hat{I}\|_2$, where $\hat{q}_{\mathrm{diff}} = \hat{q}_1^* \otimes \hat{q}_2$ measures the relative pose and $\hat{I}$ is the identity; this single norm folds translation and rotation into one scalar without the scaling factors that Euler and matrix distances require. The full reward is built in phases (maneuver, approach, handover, manipulation) with a contact-weighted term from tactile sensors, but the dual-quaternion distance is the piece the paper presents as novel and as the cause of the improved orientation behavior.

What would settle it

Re-run the same 100-episode evaluation counting only clean grasps as successes and reporting the mean over the three training seeds; the paper's own tables show that excluding indetermination drops the dual-quaternion agent to 59% success on the training prism and 69% on the short prism, and if under this stricter metric the dual-quaternion policy no longer beats the Euler-angle policy, the claimed advantage of the dual-quaternion reward for handover is not established.

Watch

Extended reading notes

Core claim

Working in simulation with a fixed giver hand holding an object and a learned receiver hand with articulated fingers, the paper argues that a dual quaternion distance $d_{\mathrm{DQ}} = \|\hat{q}_{\mathrm{diff}} - \hat{I}\|_2$, the norm of the difference between the relative pose and the identity pose, gives the reinforcement-learning agent a smoother and more accurate measure of how far the hand is from the object's grasp frame. On this metric the trained policy reduces rotational distance by about 43% over an episode, whereas an agent trained with an Euler-angle distance increases rotational distance; the dual-quaternion policy also obtains higher total success on the training prism and on a shorter prism, and its performance degrades less under a moving giver. The authors therefore claim that dual quaternions are more adequate than Euler angles or rotation matrices for computing approach rewards in $SE(3)$.

Load-bearing premise

The headline success rates count 'indetermination' episodes, cases where the simulator's collision resolution fails and the object clips through the giver hand, as successes, so the reported 91% and 94% include grasps that would not be valid in the real world.

Editorial extensions

If this is right

A reward built on dual quaternions gives a learned policy a smoother approach trajectory and a real reduction of rotation error, where an Euler-angle reward lets orientation error grow.
The dual-quaternion policy's total success exceeds the Euler policy's on the training prism and on three of four test objects, and the gap persists when the giver hand moves during handover.
Rotation-matrix-based distances, at least as implemented here, can collapse training entirely, so representation choice is not a minor implementation detail.
Robustness to unseen objects in this setup is demonstrated for geometrically similar prisms and cylinders, so the generalization claim should be read at that level of shape similarity.

Reading between the lines

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

A conservative reading of the paper's own tables is that the clean-grasp advantage of dual quaternions is smaller than the headline: excluding indetermination cases, the dual-quaternion policy succeeds in 59% of episodes on the training prism and 69% on the short prism, versus 91% and 94% total success, so part of the reported comparison rests on simulation artifacts.
Because the best of three training seeds is selected for evaluation, the reported success rates are optimistic; averaging across seeds would give a fairer estimate and could shrink the dual-quaternion versus Euler gap.
Since the policy receives the object pose directly from the simulator and no visual input, an immediate testable extension is to feed it a noisy or vision-estimated pose; the paper itself lists sim-to-real transfer as future work.
The dual-quaternion distance idea should transfer to other in-hand and regrasp tasks; a natural next experiment is to compare the same reward in a simulator with stable contact resolution, where indetermination cannot be counted as success.

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

Short version: the core comparison is probably sound, but the headline success rates lean on a metric that counts simulator collision failures as successes. I checked the tables and agree with the stress-test note. Table II: DQ is 59% strict success vs. 91% total; EULER is 17% vs. 83%. The abstract's 'best-case 94%' is 69% strict on the short prism. The authors define 'indetermination' as a physics-engine collision-resolution failure and concede it would not count as a real grasp, so the central quantitative claim as stated is not supported. The work is still worth a referee's time. It applies the group's dual quaternion reward from [7] to a new task - robot-robot handover with four-finger Allegro hands - and compares against Euler and matrix distance rewards. Full multi-finger handover learning without teleoperation is rarer than the 2-finger case, and the phase decomposition with contact-based reward shaping is clean. The strict success gap between DQ (59%) and EULER (17%) on the training object is large enough that the qualitative conclusion - DQ is a better distance metric for reward shaping here - survives the metric problem. That is the main reason I would not reject it. Soft spots, in order. First, the headline metrics: best agent of three seeds, no error bars, and Total Success used in the abstract and in Tables III and IV. Report seed-averaged strict success. Second, Section V-D evaluates the DQ agent with the same dual quaternion distance used in its reward, so the claim that DQ minimizes rotation 'better' is partly self-referential; report a common or held-out metric. Third, robustness is narrow: geometrically similar prisms and cylinders, no vision, simulation only. The conclusion is honest about these limits, but the abstract oversells. Fourth, the reward function itself is from [7], so calling it 'novel' in the abstract is an overstatement; the novelty is the application and comparison. Minor: the promised code/video link is a bare placeholder in the version I read. This is a workshop-to-conference-level paper with a fixable evaluation gap. The audience is people working on learned handover, reward shaping, or SO(3) distance metrics. I would send it to peer review with a clear request to scrutinize the metric definitions. With those revisions, it becomes a useful incremental contribution.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes an RL-based approach to dexterous object handover between two multi-finger robotic hands in simulation, using a dual-quaternion distance in the reward function to minimize rotation error. The authors train a PPO policy in IsaacLab with one prism as the training object, compare against reward functions based on Euler angles and rotation matrices, and evaluate success on seen and novel objects as well as under object-motion perturbations. They report that the dual-quaternion agent achieves 91% total success on the training prism, 94% in the best case on a short prism, and small degradation under perturbations, and they argue that the dual-quaternion representation minimizes rotation distance more effectively than Euler angles.

Significance. If the central claims can be properly supported, the work is a useful empirical contribution to RL-based dexterous manipulation: it extends dual-quaternion rewards from single-arm approach to two-hand object handover, provides a comparison with two alternative rotation representations, and releases code and videos. The perturbation study and the explicit limitations section are also valuable. However, the significance currently hinges on success metrics that count a simulator artifact as success and on a distance-minimization analysis that compares each policy with its own reward metric; these issues must be resolved before the stated performance numbers can be taken at face value.

major comments (4)

[Section V-C, Tables II-IV, Abstract] The primary success metric, 'Total Success', counts 'Indetermination' episodes as successes, where Indetermination is explicitly defined as a simulator collision-resolution failure in which the object clips through the UR5e hand and would not count as a real-world grasp. This inflates the headline numbers: Table II reports the DQ agent as 59% Succ + 32% Ind = 91% Total and the EULER agent as 17% Succ + 66% Ind = 83% Total; Table III's short-prism row gives DQ 69% Succ + 25% Ind = 94% Total, which is the '94% best-case scenario' cited in the abstract; Table IV's perturbation results are also Total-Success based. Because the paper's central claim is that the policy 'successfully perform[s]' the handover, the valid metric is strict success. The authors should re-report all results using strict success (and failure) rates, update the abstract and conclusions to reflect the corrected numbers, and either remove Indetermination from the success definition or clearly present it as a separate non-success category.
[Section V-D, Fig. 7] The distance-minimization comparison is circular. The DQ agent's trajectories are evaluated using the dual-quaternion distance in Eq. (9), while the EULER agent's trajectories are evaluated using the Euler distance in Eq. (10); in each case, the metric is the one used to shape that agent's reward. The statement that the DQ agent 'minimized the rotational poses from the starting pose around 43% on average' while the EULER agent 'increased it' therefore does not establish that the DQ reward yields better rotation alignment; it may only show that each policy minimizes its own training objective. To support the claimed superiority of dual-quaternion rewards, both policies should be evaluated with a single, externally defined rotation metric (for example, the SO(3) geodesic angle or the rotation part of the dual-quaternion difference applied to both agents), and the 43% figure should be recomputed under that common metric.
[Section V-C] The evaluation selects 'the best agents for each case' from three training seeds, and no per-seed results or variance are reported. All tables and the abstract's 94% figure are based on this best-seed selection. Because the central claim is that the DQ reward outperforms the EULER reward, the comparison should be robust to seed variability. The authors should report mean and standard deviation across seeds, or at least provide the per-seed strict-success rates for both agents, so that the reader can assess whether the advantage is consistent rather than a single-lucky-seed artifact.
[Section V-C, Conclusion] The generalization claim is narrower than the abstract suggests. The out-of-distribution objects are only a shorter prism and two cylinders, all geometrically similar to the training prism, and the policy receives ground-truth object pose from the simulator with no visual perception, as the conclusion acknowledges. The abstract's phrase 'robustness of our policy with novel objects' should be qualified to 'novel object geometries in simulation with oracle state', and the limitations on visual perception and physics-engine artifacts should be surfaced in the abstract or introduction rather than appearing only in the final section.

minor comments (6)

[Table I] The 'Magnitude' operation is written as an equality rather than a norm; please clarify that the magnitude of a dual quaternion is computed as the square root of the product with its conjugate, or otherwise state the intended definition.
[Section IV-C] Eq. (10) uses scaling weights ψ and μ, while Eq. (12) uses ψ and β; since the numerical values are identical (2.1 and 0.32), please use consistent notation to avoid confusion.
[Fig. 5] The training curves show only the mean reward with no error bars or per-seed shading; adding seed variance would make the learning comparison more informative.
[Abstract and Section V-C] The phrase 'after 100 experiments' should specify that these are 100 evaluation episodes per condition, and the abstract should state that the reported rates are 'Total Success' as defined in Section V-C, since that term is non-standard.
[Throughout] There are several typographical and grammatical errors, including 'inS0(3)' in the introduction (should be SO(3)), 'in contrast of using' (should be 'in contrast to using'), and 'the robot is approaching better the grasp frame' in Section V-C.
[References] Reference [12] lists the arXiv identifier 1701.07274 but a DOI corresponding to 2412.05265; please correct the citation so that the identifier and DOI match.

Circularity Check

1 steps flagged · score 4.0 of 10

Rotation-minimization evidence in Section V-D is self-referential, but headline success-rate comparisons remain externally defined.

self definitional [Section V-D (Distance minimization with dual quaternions), using Eq. (9); reward defined in Section IV-C-1 and Eq. (3)/(8)]
"we calculate the distance between frames using dual quaternions using (9) as presented in [16]. In this work, we propose this reward function to perform the object handover task. ... For the DQ agent, we calculated the mean distances following (9) in each step until the target pose is reached."

The DQ agent was trained with a reward whose core term is d_DQ from Eq. (9), via r_BASE = eta * e^{-d_t}. Section V-D then claims that the DQ agent 'minimized the rotational poses ... around 43%' by measuring the DQ agent with Eq. (9) itself. That is an evaluation of the agent on its own training objective, so the reduction in d_DQ is a restatement of the fitted reward, not an independent finding. The EULER agent, by contrast, was trained on the different weighted objective of Eq. (10), and its rotation component is extracted from that same objective. Comparing the two under different metrics therefore makes the conclusion that dual quaternions are more adequate partly true by construction.

full rationale

The only concrete circularity I can exhibit is in Section V-D: the DQ agent's reward is built from the dual-quaternion distance of Eq. (9), and the paper then measures that same distance to show the DQ agent minimizes rotation. This is a self-referential supporting experiment, not an independent test of the reward representation. The headline success-rate claims in Tables II-IV are defined against external task outcomes (grasp, no fall) and are therefore not circular, even though the 'Total Success' metric counts simulator indetermination as success; that is a metric-validity concern rather than a derivational circularity. Citation [7] is prior work by one of the authors, but the reward function is fully specified in the present paper and is not justified by the citation alone, so it is not load-bearing. Because the main handover-success comparison does not reduce to the reward definition, the overall circularity is limited to the rotation-distance sub-claim, giving a moderate score of 4 rather than a higher one.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or forces. Its load-bearing assumptions are: the dual quaternion distance metric is appropriate; the giver is static during training; ground-truth object pose is available; and the hand-tuned reward phases are effective. The free parameters are all hand-selected scalars in the reward and distance equations, plus the perturbation velocities. These are not fitted to outcome data in a statistical sense, but they are tuned to make the training work, so they count against the paper's claims of a clean comparison.

free parameters (7)

eta_0 = 1
Base weight in the basic approach reward r_BASE (Eq. 3), chosen by hand.
alpha = 12
Weight for the manipulation phase reward (Eq. 7), chosen by hand.
psi = 2.1
Translation scaling factor in the EULER and MATRIX distance equations (Eq. 10 and 12), hand-tuned to make translation and rotation distances comparable in magnitude.
mu = 0.32
Rotation scaling factor in the EULER distance (Eq. 10), hand-tuned.
beta = 0.32
Rotation scaling factor in the MATRIX distance (Eq. 12), hand-tuned.
perturbation velocities = 0.03 m/s linear, 0.16 rad/s angular
Chosen for the moving-giver perturbation experiments (Section V-E); the paper states these are 40% lower than the receiving robot's velocity, but the exact values are arbitrary choices.
contact weight vector w_c = not fully specified
Weights for the boolean contact sensors that shape the reward in Eq. (8). The text says lower phalanges are more important than tips, but the exact values are not reported.

assumptions (4)

standard math Dual quaternion algebra operations (Table I) provide a correct and valid metric for pose differences in SE(3).
The reward and evaluation distances rely on the dual quaternion difference and the identity comparison in Eq. (9), using the algebra introduced in the references [14], [15].
domain assumption The giver robot (UR5e) holds the object and remains stationary during training; it moves only in perturbation tests.
Section IV-A states the UR5e 'holds the object without moving during the whole episode' and 'mimics a human giver which should not be trained.' This is a simplification that makes the task single-agent.
domain assumption Ground-truth object pose is available to the policy as an observation; no visual perception is used.
The policy observations include 'the Cartesian poses of the GEN3 and the object' (Section IV-A), and the conclusion explicitly lists the lack of visual perception as a limitation.
domain assumption The hand-crafted phase-based reward shaping (maneuver, approach, handover, manipulation) is necessary and sufficient to learn the task.
The reward architecture in Eq. (8) with modifiers (4)-(7) is manually designed. The paper does not provide an ablation or a derivation showing these phases are optimal or necessary.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Dexterous Object Handover." pith.science (2026). https://pith.science/paper/XYBT2ANJ

@misc{pith2026250616822,
  author       = {Pith},
  title        = {Pith review of: Learning Dexterous Object Handover},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYBT2ANJ}},
  note         = {Machine review of arXiv:2506.16822}
}

read the original abstract

Object handover is an important skill that we use daily when interacting with other humans. To deploy robots in collaborative setting, like houses, being able to receive and handing over objects safely and efficiently becomes a crucial skill. In this work, we demonstrate the use of Reinforcement Learning (RL) for dexterous object handover between two multi-finger hands. Key to this task is the use of a novel reward function based on dual quaternions to minimize the rotation distance, which outperforms other rotation representations such as Euler and rotation matrices. The robustness of the trained policy is experimentally evaluated by testing w.r.t. objects that are not included in the training distribution, and perturbations during the handover process. The results demonstrate that the trained policy successfully perform this task, achieving a total success rate of 94% in the best-case scenario after 100 experiments, thereby showing the robustness of our policy with novel objects. In addition, the best-case performance of the policy decreases by only 13.8% when the other robot moves during the handover, proving that our policy is also robust to this type of perturbation, which is common in real-world object handovers.

Figures

Figures reproduced from arXiv: 2506.16822 by the authors.

**Figure 2.** Overview of the RL environment used in this work, comprising a single PPO agent. The observations consist of the Euclidean translations and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 1.** In each stage, we change a simple initial reward [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Placement of the contact sensors on the Allegro hand. Each green [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Average reward for all the agents trained for the object handover [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 2.** Figure 2: The single object training aims to reduce computa [PITH_FULL_IMAGE:figures/full_fig_p005_2.png]

**Figure 6.** Figure 6: Graphical description of the different cases contemplated during the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Distance minimization of the agents for all selected objects. The DQ agent showed a greater minimization of the rotation distance than the EULER [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 12 canonical work pages

[7]

Robotic approach trajectory using reinforcement learning with dual quaternions,

D. Frau-Alfaro, S. T. Puente, I. De Loyola P ´aez-Ubieta, and E. Velasco-S´anchez, “Robotic approach trajectory using reinforcement learning with dual quaternions,” in7th Iberian Robotics Conference (ROBOT), 2024. doi: 10.1109/ROBOT61475.2024.10796878

work page arXiv 2024
[1]

Object handovers: A review for robotics,

V . Ortenzi, A. Cosgun, T. Pardi, W. P. Chan, E. Croft, and D. Kuli ´c, “Object handovers: A review for robotics,”IEEE Trans- actions on Robotics, vol. 37, no. 6, pp. 1855–1873, 2021. doi: 10.1109/TRO.2021.3075365

arXiv 2021
[2]

Human–robot object handover: Recent progress and future direction,

D. Haonan, Y . Yifan, L. Daheng, and W. Peng, “Human–robot object handover: Recent progress and future direction,”Biomimetic Intelligence and Robotics, vol. 4, no. 1, p. 100145, 2024. doi: 10.1016/j.birob.2024.100145

arXiv 2024
[3]

Handover control for human-robot and robot-robot collaboration,

M. Costanzo, G. D. Maria, and C. Natale, “Handover control for human-robot and robot-robot collaboration,”Frontiers in Robotics and AI, vol. 8, 2021. doi: 10.3389/frobt.2021.672995

work page arXiv 2021
[4]

Efficient bimanual han- dover and rearrangement via symmetry-aware actor-critic learning,

Y . Li, C. Pan, H. Xu, X. Wang, and Y . Wu, “Efficient bimanual han- dover and rearrangement via symmetry-aware actor-critic learning,” in IEEE International Conference on Robotics and Automation (ICRA),

work page
[5]

3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing,

B. Huang, Y . Wang, X. Yang, Y . Luo, and Y . Li, “3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing,” in8th Annual Conference on Robot Learning, 2024. doi: 10.48550/arXiv.2410.24091

arXiv doi:10.48550/arxiv.2410.24091 2024
[6]

Genh2r: Learn- ing generalizable human-to-robot handover via scalable simulation, demonstration, and imitation,

Z. Wang, J. Chen, Z. Chen, P. Xie, R. Chen, and L. Yi, “Genh2r: Learn- ing generalizable human-to-robot handover via scalable simulation, demonstration, and imitation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 362–16 372. doi: 10.1109/CVPR52733.2024.01548

arXiv 2024
[8]

Bidirectional human–robot bimanual handover of big planar object with vertical posture,

W. He, J. Li, Z. Yan, and F. Chen, “Bidirectional human–robot bimanual handover of big planar object with vertical posture,”IEEE Transactions on Automation Science and Engineering, pp. 1180–1191,

work page

Show all 23 references

[9]

Naturalistic robot-to-human bimanual handover in complex environments through multi-sensor fusion,

S. E. Ovur and Y . Demiris, “Naturalistic robot-to-human bimanual handover in complex environments through multi-sensor fusion,”IEEE Transactions on Automation Science and Engineering, pp. 3730–3741,
[10]

Dynamic handover: Throw and catch with bimanual hands,

B. Huang, Y . Chen, T. Wang, Y . Qin, Y . Yang, N. Atanasov, and X. Wang, “Dynamic handover: Throw and catch with bimanual hands,” in7th Annual Conference on Robot Learning, 2023. doi: 10.48550/arXiv.2309.05655

2023 doi arXiv
[11]

Sim-to- real reinforcement learning for vision-based dexterous manipula- tion on humanoids,

T. Lin, K. Sachdev, L. Fan, J. Malik, and Y . Zhu, “Sim-to- real reinforcement learning for vision-based dexterous manipula- tion on humanoids,”arXiv preprint arXiv:2502.20396, 2025. doi: 10.48550/arXiv.2502.20396

2025 doi arXiv
[12]

Deep reinforcement learning: An overview,

Y . Li, “Deep reinforcement learning: An overview,”arXiv preprint arXiv:1701.07274, 2017. doi: 10.48550/arXiv.2412.05265

2017 doi arXiv
[13]

Reinforcement learning,

M. A. Wiering and M. Van Otterlo, “Reinforcement learning,”Adap- tation, learning, and optimization, vol. 12, no. 3, p. 729, 2012

2012
[14]

Dual quaternions,

Y .-B. Jia, “Dual quaternions,”Iowa State University: Ames, IA, USA, 2013. [Online]. Available: https://faculty.sites.iastate.edu/jia/ files/inline-files/dual-quaternion.pdf

2013
[15]

Approaching dual quaternions from matrix algebra,

F. Thomas, “Approaching dual quaternions from matrix algebra,”IEEE Transactions on Robotics, 2014. doi: 10.1109/TRO.2014.2341312

2014
[16]

Dualquat-loam: Lidar odometry and mapping parameterized on dual quaternions,

E. P. Velasco-S ´anchez, L. F. Recalde, G. Li, F. A. Candelas-Herias, S. T. Puente-Mendez, and F. Torres-Medina, “Dualquat-loam: Lidar odometry and mapping parameterized on dual quaternions,”Robotics and Autonomous Systems, 2025. doi: 10.1016/j.robot.2025.105009

2025
[17]

Learning positioning policies for mobile manipulation operations with deep reinforcement learning,

A. Iriondo, E. Lazkano, A. Ansuategi, A. Rivera, I. Lluvia, and C. Tub ´ıo, “Learning positioning policies for mobile manipulation operations with deep reinforcement learning,”International journal of machine learning and cybernetics, 2023. doi: 10.1007/s13042-023- 01815-8

2023 doi
[18]

Orbit: A unified simulation framework for interactive robot learning environments,

M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg, “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE Robotics and Automa- tion Letter...

2023
[19]

Stable-baselines3: Reliable reinforcement learning implementations,

A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-baselines3: Reliable reinforcement learning implementations,”Journal of Machine Learning Research. [Online]. Available: http://jmlr.org/papers/v22/20-1364.html
[20]

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. doi: 10.48550/arXiv.1707.06347

2017 doi arXiv
[2022]

doi: 10.1109/TASE.2020.3043480

2020
[2023]

doi: 10.1109/ICRA48891.2023.10160739

2023
[2024]

doi: 10.1109/TASE.2023.3284668

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [7]

Robotic approach trajectory using reinforcement learning with dual quaternions,

D. Frau-Alfaro, S. T. Puente, I. De Loyola P ´aez-Ubieta, and E. Velasco-S´anchez, “Robotic approach trajectory using reinforcement learning with dual quaternions,” in7th Iberian Robotics Conference (ROBOT), 2024. doi: 10.1109/ROBOT61475.2024.10796878

work page arXiv 2024

[2] [1]

Object handovers: A review for robotics,

V . Ortenzi, A. Cosgun, T. Pardi, W. P. Chan, E. Croft, and D. Kuli ´c, “Object handovers: A review for robotics,”IEEE Trans- actions on Robotics, vol. 37, no. 6, pp. 1855–1873, 2021. doi: 10.1109/TRO.2021.3075365

arXiv 2021

[3] [2]

Human–robot object handover: Recent progress and future direction,

D. Haonan, Y . Yifan, L. Daheng, and W. Peng, “Human–robot object handover: Recent progress and future direction,”Biomimetic Intelligence and Robotics, vol. 4, no. 1, p. 100145, 2024. doi: 10.1016/j.birob.2024.100145

arXiv 2024

[4] [3]

Handover control for human-robot and robot-robot collaboration,

M. Costanzo, G. D. Maria, and C. Natale, “Handover control for human-robot and robot-robot collaboration,”Frontiers in Robotics and AI, vol. 8, 2021. doi: 10.3389/frobt.2021.672995

work page arXiv 2021

[5] [4]

Efficient bimanual han- dover and rearrangement via symmetry-aware actor-critic learning,

Y . Li, C. Pan, H. Xu, X. Wang, and Y . Wu, “Efficient bimanual han- dover and rearrangement via symmetry-aware actor-critic learning,” in IEEE International Conference on Robotics and Automation (ICRA),

work page

[6] [5]

3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing,

B. Huang, Y . Wang, X. Yang, Y . Luo, and Y . Li, “3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing,” in8th Annual Conference on Robot Learning, 2024. doi: 10.48550/arXiv.2410.24091

arXiv doi:10.48550/arxiv.2410.24091 2024

[7] [6]

Genh2r: Learn- ing generalizable human-to-robot handover via scalable simulation, demonstration, and imitation,

Z. Wang, J. Chen, Z. Chen, P. Xie, R. Chen, and L. Yi, “Genh2r: Learn- ing generalizable human-to-robot handover via scalable simulation, demonstration, and imitation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 362–16 372. doi: 10.1109/CVPR52733.2024.01548

arXiv 2024

[8] [8]

Bidirectional human–robot bimanual handover of big planar object with vertical posture,

W. He, J. Li, Z. Yan, and F. Chen, “Bidirectional human–robot bimanual handover of big planar object with vertical posture,”IEEE Transactions on Automation Science and Engineering, pp. 1180–1191,

work page