Pith. sign in

REVIEW 3 major objections 5 minor 40 references

A robot arm and hand can catch human-thrown objects using only pixel-level visual cues from a single RGB camera, without ever estimating the object's 3D position.

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

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

T0 review · deepseek-v4-flash

2026-08-02 20:36 UTC pith:CAKQV4O5

load-bearing objection RGB-only catching works in zero-shot sim-to-real, but the SAM2 perception link is under-measured and the real-world stats are thin. the 3 major comments →

arxiv 2602.22733 v1 pith:CAKQV4O5 submitted 2026-02-26 cs.RO

Pixel2Catch: Multi-Agent Sim-to-Real Transfer for Agile Manipulation with a Single RGB Camera

Seongyong Kim , Junhyeon Cho , Kang-Won Lee , Soo-Chul Lim This is my paper
classification cs.RO
keywords pixel-level featuressim-to-real transfermulti-agent reinforcement learningrobotic catchingdexterous manipulationRGB-only perceptiondomain randomization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to prove that dynamic catching does not require explicit 3D object position estimation. The authors train a two-part policy — one agent moves the arm, one closes the fingers — in simulation using multi-agent reinforcement learning, feeding the policy a six-number vector of pixel-level visual cues (object center, width, height, and their changes over time) extracted from a single RGB camera. They then transfer the policy to a real UR5e arm with an Allegro hand, with no real-world fine-tuning, and report catch success rates of 43–63% across three thrown objects. If right, this shows that cheap monocular perception plus role-split learning can replace depth sensors and motion-capture systems in agile manipulation.

Core claim

The paper's central claim is that a six-dimensional pixel-feature vector — object center coordinates, width, height, and their temporal differences — is sufficient input for a catching policy, and that splitting the policy into an arm agent and a hand agent with role-specific rewards makes this learnable and transferable. On a UR5e arm with an Allegro hand, the simulation-trained policies catch human-thrown objects at 43-63% success in the real world, outperforming baselines that use only single visual cues or a single combined policy. The authors conclude that effective dynamic catching can be realized without explicit 3D object position estimation, relying solely on relative visual cues fr

What carries the argument

The key mechanism is the pixel-level feature vector z_pixel = (c_x, c_y, Δc_x, Δc_y, Δw, Δh), extracted from a segmentation-derived bounding box in the RGB image. Center coordinates give lateral and vertical motion; width and height encode distance via scale; the temporal differences (Δ) provide velocity. This compact, task-relevant representation replaces explicit 3D state estimation. The second mechanism is heterogeneous multi-agent reinforcement learning: the arm and hand are independent agents with separate observations and rewards, trained jointly with centralized value functions. The authors argue this role split stabilizes learning for the high-degree-of-freedom system and improves si

Load-bearing premise

The whole perception chain rests on the assumption that the segmentation model will reliably locate the object in every 30 Hz camera frame; the paper never measures its latency or failure rate, and if segmentation lags or drifts, the pixel features — and the policy — are corrupted.

What would settle it

Run the trained policy with the segmentation model's inference time artificially increased to just over one control cycle (33 ms) and observe whether catch rate collapses; or verify whether the pixel-feature policy still catches when the object is thrown directly toward the camera so that its image-space center barely moves and only scale changes.

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

If this is right

  • Monocular RGB catching removes the need for depth sensors, motion capture, or explicit 3D tracking, lowering cost and setup complexity.
  • Because both center and scale cues are needed, future visual-policy designs should combine positional and scale-change inputs rather than relying on one.
  • Role-split multi-agent training may be a general recipe for high-DoF manipulation beyond catching, such as any arm-plus-hand task.
  • The zero-shot sim-to-real transfer indicates that pixel-feature representations are more domain-invariant than metric 3D estimates.
  • The 43-63% real-world success rate provides a baseline for future RGB-only catching systems to measure against.

Where Pith is reading between the lines

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

  • The paper never measures the segmentation model's latency or failure rate; if a faster or more reliable segmenter were available, the same policy could plausibly handle faster throws or cluttered backgrounds.
  • Because the policy sees only pixel features, it should be insensitive to camera calibration and robot-camera pose errors — a testable advantage over 3D-estimation approaches under camera misalignment.
  • Failures are likely concentrated in cases where scale changes are ambiguous (object thrown straight at the camera) or where segmentation momentarily fails; analyzing these cases could guide targeted improvements.
  • The same template could extend to bimanual catching (the paper's stated future work) simply by adding a second arm agent with its own role reward.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Pixel2Catch, a framework for catching human-thrown objects with a UR5e arm and Allegro hand using a single fixed RGB camera. Object motion is represented by a six-dimensional pixel-feature vector z_pixel = {c_x, c_y, Δc_x, Δc_y, Δw, Δh} extracted from SAM2 segmentation of consecutive frames (Eq. (1)). Two policies, one for the arm and one for the hand, are trained with MAPPO using role-specific observations and rewards in Isaac Lab, with system identification and domain randomization, and deployed zero-shot. Simulation evaluations on seen and unseen objects report tracking and success rates; real-world evaluations on three objects report tracking rates of 70–73% and success rates of 43–63%, outperforming the baselines. The paper concludes that effective catching is achievable without explicit 3D object position estimation.

Significance. If the claims hold, this is a useful existence proof that a compact pixel-space representation plus a heterogeneous MARL decomposition can transfer to real hardware for a high-DoF dynamic manipulation task. The paper has clear strengths: simulation results are averaged over three seeds with standard deviations; ablations isolate the contribution of center vs. width/height cues; the real-world evaluation is zero-shot and no parameters are fitted to the real-world trial outcomes; system identification and domain randomization are described; and the architecture details are enough to understand the method. The main risk is the unverified perception pipeline: the only object-motion signal is z_pixel, and its real-world availability and quality at 30 Hz are assumed, not measured. The central claim is defensible but currently supported by evidence that needs to be strengthened in specific places.

major comments (3)
  1. [Sec. IV-B, V-D, Eq. (1)] The central perceptual claim rests on SAM2 producing the six pixel features reliably at 30 Hz, but no measurements or failure analysis are provided. The paper does not report SAM2 inference latency on the deployment hardware, segmentation accuracy, per-frame failure rate, or how the mask is initialized/prompted for a human-thrown object (first-frame box, automatic mask generation, or an external detector). It also does not state whether the simulation z_pixel is computed from ground-truth masks or from rendered RGB through the same segmentation. Since Table I adds only 5-pixel corner perturbation, the sim-to-real distribution shift in z_pixel is unquantified. If SAM2 lags or drifts, the policy receives stale or wrong cues regardless of how well the RL policies are trained. Please measure latency and accuracy, specify the initialization mechanism, and characterize the observed failure mod
  2. [Table III / Sec. V-D] The real-world comparative claims are based on 30 trials per object with no confidence intervals, error bars, or significance tests. For example, proposed vs. S-A RL success rates are Cube 63 vs. 33, L-block 43 vs. 20, Triangle 43 vs. 20. With n=30, an exact binomial 95% CI for 43% is roughly [25, 63]%, so some differences (e.g., Triangle 43 vs. 20) are within the noise range. The zero results for only-WH should also be accompanied by confidence intervals (0/30 gives an upper bound near 11%). The authors should report binomial CIs and/or a paired test across repeated trial blocks before claiming consistent superiority over baselines.
  3. [Sec. IV-F and V-D] The sim-to-real argument depends on domain randomization and system identification, but the paper does not quantify how close simulated object trajectories and robot tracking are to real ones, nor does it evaluate the policy's sensitivity to the actual SAM2 output errors. The 5-pixel augmentation is not justified relative to observed segmentation error; real segmentation errors are correlated and can be much larger for fast-moving objects. Please provide, at minimum, a characterization of real-world z_pixel error against manual labels (or another reference), and show that the trained policy is robust to errors at that scale. Without this, the claimed transfer robustness is not fully supported.
minor comments (5)
  1. [Sec. II-B] Typo: 'UA V swarm control' should be 'UAV swarm control'.
  2. [Sec. IV-E] The training hyperparameter list is incomplete for reproducibility: learning rate, batch size, GAE lambda, number of epochs, and total environment steps are not reported, even though the MAPPO/skrl implementation is cited.
  3. [Table III] The table caption says 'averaged over 30 trials'; these are proportions, not averages. Please report raw counts (e.g., 19/30) and exact binomial confidence intervals.
  4. [Sec. V and Table III] The w/o-PF baseline is included in simulation but not evaluated in the real-world experiments. Please state why, or include it if feasible, so the comparison is uniform.
  5. [Sec. III-B] The exact positions of the three fixed Allegro joints and the camera pose (extrinsic calibration) are not given; these details are useful for reproduction.

Circularity Check

0 steps flagged

No significant circularity: real-world evaluation is a zero-shot holdout and no prediction reduces to a fitted input.

full rationale

The derivation chain is self-contained and externally evaluated. The policies are trained in simulation with rewards and observations defined in Secs. IV-C/IV-D, then deployed directly on the real robot with no fine-tuning: Sec. V-D states 'all policies are transferred directly from simulation to the real robot without any additional real-world fine-tuning.' The real-world success rates in Table III are holdout measurements; no parameter, reward weight, or domain-randomization range is fitted to those trials, so the reported 43–63% success is not forced by construction. Eq. (1) defines z_pixel from segmentation-based center, width, height, and their temporal differences; this is an input representation for RL, not a fitted output. The central claim—that catching is possible without explicit 3D position—is supported by comparisons against only-Center, only-WH, and S-A RL baselines in both simulation and real-world tables. The self-citations [18] and [20] appear only in the related-work enumeration of static manipulation settings and do not carry any load-bearing premise; there is no imported uniqueness theorem, no ansatz smuggled via self-citation, and no renamed known result. The lack of reported SAM2 inference latency and segmentation accuracy (Secs. IV-B, V-D) is a real correctness risk for the 30 Hz real-time claim, but it concerns an unverified external tool assumption, not circularity: the perception chain is not equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim depends on hand-selected reward weights, perturbation bounds, domain-randomization ranges, and optimized dynamics parameters, plus unproven assumptions about SAM2 reliability, simulation fidelity, and the sufficiency of pixel cues. No new physical entities are introduced.

free parameters (5)
  • Reward coefficients (arm and hand) = lambda_succ=10.0, lambda_fail=5.0, lambda_app=0.1, lambda_act=0.01, r_time=-0.01
    Hand-chosen weights in Eqs. (2)-(3); balance reaching, grasping, and penalties. Not fitted to real-world benchmark data.
  • Pixel perturbation bound = up to 5 pixels on object-region corners
    Hand-picked augmentation in Sec. IV-B to make features robust to segmentation uncertainty.
  • Domain randomization ranges = joint stiffness [0.7,1.3], joint damping [0.7,1.3], action noise sigma=0.02/0.03, obs noise sigma=0.005, initial joint p
    Hand-selected distributions in Table I; these directly affect the sim-to-real transfer claim.
  • System-identification parameters = optimized joint actuation gains, damping, friction, armature
    Fitted to real-robot trajectory data in Sec. IV-F; central to zero-shot transfer.
  • MAPPO/network hyperparameters = gamma=0.99, clip eps=0.2, KL threshold=0.016, hidden [512,256,128], ELU activation, 3 seeds
    Chosen RL hyperparameters in Sec. IV-E; no sensitivity analysis is reported.
axioms (5)
  • domain assumption MAPPO with a centralized value function and decentralized policies learns a cooperative arm/hand catching policy in this high-DoF setting.
    No convergence proof is given; relies on ref [11] and empirical practice.
  • domain assumption The pixel-feature vector {cx, cy, delta_cx, delta_cy, delta_w, delta_h} is sufficient to infer the object's 3D motion needed for catching.
    Core representation claim of Sec. IV-B; argued from human perception, not proven.
  • domain assumption Isaac Lab simulation, after system identification and domain randomization, is a faithful enough proxy for real robot dynamics and throw distributions.
    Secs. III-C and IV-F; required for zero-shot transfer. The paper itself notes unmodeled real-world effects like object deformation and aerodynamics.
  • domain assumption SAM2 provides reliable object segmentation in diverse real-world conditions at the 30 Hz control rate.
    Sec. IV-B; no latency, accuracy, or failure-rate evaluation is reported.
  • domain assumption The hand reward, averaged over palm and fingertip distance improvements, induces stable grasping rather than merely approaching the object.
    Eq. (3); reward shaping is assumed to produce the intended behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 10571 in / 15374 out tokens · 154342 ms · 2026-08-02T20:36:56.385642+00:00 · methodology

0 comments
read the original abstract

To catch a thrown object, a robot must be able to perceive the object's motion and generate control actions in a timely manner. Rather than explicitly estimating the object's 3D position, this work focuses on a novel approach that recognizes object motion using pixel-level visual information extracted from a single RGB image. Such visual cues capture changes in the object's position and scale, allowing the policy to reason about the object's motion. Furthermore, to achieve stable learning in a high-DoF system composed of a robot arm equipped with a multi-fingered hand, we design a heterogeneous multi-agent reinforcement learning framework that defines the arm and hand as independent agents with distinct roles. Each agent is trained cooperatively using role-specific observations and rewards, and the learned policies are successfully transferred from simulation to the real world.

Figures

Figures reproduced from arXiv: 2602.22733 by Junhyeon Cho, Kang-Won Lee, Seongyong Kim, Soo-Chul Lim.

Figure 1
Figure 1. Figure 1: We propose Pixel2Catch, an RGB-only robotic catching system without explicit 3D position estimation. The system consists of a robot arm equipped with a multi-fingered hand and a single RGB camera. Inspired by human visual perception, object motion is inferred from pixel-level features in image space rather than metric 3D coordinates. Policies trained in simulation are transferred directly to the real robot… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of the system and experimental setup. Each policy ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Objects used for training (top), validation (mid [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of pixel-level features in simulation [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Tracking and success rates over training. Results are [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the results of deploying a trained policy in real-world experiments. This figure presents real-world [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

40 extracted references · 11 linked inside Pith

  1. [1]

    Catching objects in flight,

    S. Kim, A. Shukla, and A. Billard, “Catching objects in flight,”IEEE Transactions on Robotics, vol. 30, no. 5, pp. 1049–1065, 2014

    2014

  2. [2]

    Cushioncatch: A compliant catching mechanism for mobile manipulators via combined optimization and learning,

    B. Chen, K. Fan, Q. Yang, Y . Cheng, H. Liu, K. Dong, C. Xia, L. Han, and B. Liang, “Cushioncatch: A compliant catching mechanism for mobile manipulators via combined optimization and learning,”arXiv preprint arXiv:2409.14754, 2024

    arXiv 2024

  3. [3]

    A dynamical system approach for softly catching a flying object: Theory and experiment,

    S. S. M. Salehian, M. Khoramshahi, and A. Billard, “A dynamical system approach for softly catching a flying object: Theory and experiment,”IEEE Transactions on Robotics, vol. 32, no. 2, pp. 462– 471, 2016

    2016

  4. [4]

    Ima-catcher: An impact-aware nonprehensile catching framework based on combined optimization and learning,

    F. Tassi, J. Zhao, G. J. Lahr, L. Gava, M. Monforte, A. Glover, C. Bartolozzi, and A. Ajoudani, “Ima-catcher: An impact-aware nonprehensile catching framework based on combined optimization and learning,”The International Journal of Robotics Research, p. 02783649251345851, 2025

    2025

  5. [5]

    Dexcatch: Learning to catch arbitrary objects with dexterous hands,

    F. Lan, S. Wang, Y . Zhang, H. Xu, O. Oseni, Z. Zhang, Y . Gao, and T. Zhang, “Dexcatch: Learning to catch arbitrary objects with dexterous hands,”arXiv preprint arXiv:2310.08809, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Catch it! learn- ing to catch in flight with mobile dexterous hands,

    Y . Zhang, T. Liang, Z. Chen, Y . Ze, and H. Xu, “Catch it! learn- ing to catch in flight with mobile dexterous hands,”arXiv preprint arXiv:2409.10319, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    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,” inConference on Robot Learning. PMLR, 2023, pp. 1887–1902

    2023

  8. [8]

    Deep reinforcement learning based moving object grasping,

    P. Chen and W. Lu, “Deep reinforcement learning based moving object grasping,”Information Sciences, vol. 565, pp. 62–76, 2021

    2021

  9. [9]

    Learning dexterous bimanual catch skills through adversarial-cooperative heterogeneous-agent reinforce- ment learning,

    T. Kim, Y . Yoon, and J. Kim, “Learning dexterous bimanual catch skills through adversarial-cooperative heterogeneous-agent reinforce- ment learning,”arXiv preprint arXiv:2502.11437, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafsonet al., “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    The surprising effectiveness of ppo in cooperative multi-agent games,

    C. Yu, A. Velu, E. Vinitsky, J. Gao, Y . Wang, A. Bayen, and Y . Wu, “The surprising effectiveness of ppo in cooperative multi-agent games,”Advances in Neural Information Processing Systems, vol. 35, pp. 24 611–24 624, 2022

    2022

  12. [12]

    Real-time scheduling for flexible job shop with agvs using multiagent reinforce- ment learning and efficient action decoding,

    Y . Li, Q. Wang, X. Li, L. Gao, L. Fu, Y . Yu, and W. Zhou, “Real-time scheduling for flexible job shop with agvs using multiagent reinforce- ment learning and efficient action decoding,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2025

    2025

  13. [13]

    Language-driven policy distillation for cooperative driving in multi- agent reinforcement learning,

    J. Liu, C. Xu, P. Hang, J. Sun, M. Ding, W. Zhan, and M. Tomizuka, “Language-driven policy distillation for cooperative driving in multi- agent reinforcement learning,”IEEE Robotics and Automation Letters, 2025

    2025

  14. [14]

    Distributed reinforce- ment learning for cooperative multi-robot object manipulation,

    G. Ding, J. J. Koh, K. Merckaert, B. Vanderborght, M. M. Nicotra, C. Heckman, A. Roncone, and L. Chen, “Distributed reinforce- ment learning for cooperative multi-robot object manipulation,”arXiv preprint arXiv:2003.09540, 2020

    Pith/arXiv arXiv 2003

  15. [15]

    Safe multi-agent reinforcement learning for bimanual dexterous manipulation,

    W. Zhan and P. Chin, “Safe multi-agent reinforcement learning for bimanual dexterous manipulation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 12 420–12 427

    2024

  16. [16]

    Rotating without seeing: Towards in-hand dexterity through touch,

    Z.-H. Yin, B. Huang, Y . Qin, Q. Chen, and X. Wang, “Rotating without seeing: Towards in-hand dexterity through touch,”arXiv preprint arXiv:2303.10880, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Lessons from learning to spin “pens

    J. Wang, Y . Yuan, H. Che, H. Qi, Y . Ma, J. Malik, and X. Wang, “Lessons from learning to spin “pens”,”CoRL, 2024

    2024

  18. [18]

    Dextouch: Learning to seek and manipulate objects with tactile dexterity,

    K.-W. Lee, Y . Qin, X. Wang, and S.-C. Lim, “Dextouch: Learning to seek and manipulate objects with tactile dexterity,”IEEE Robotics and Automation Letters, 2024

    2024

  19. [19]

    Learning visuotactile skills with two multifingered hands,

    T. Lin, Y . Zhang, Q. Li, H. Qi, B. Yi, S. Levine, and J. Malik, “Learning visuotactile skills with two multifingered hands,”arXiv preprint arXiv:2404.16823, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Progressive policy learning: A hierarchical framework for dexterous bimanual manipula- tion,

    K.-W. Lee, J.-W. Lee, S. Kim, and S.-C. Lim, “Progressive policy learning: A hierarchical framework for dexterous bimanual manipula- tion,”Mathematics, vol. 13, no. 22, p. 3585, 2025

    2025

  21. [21]

    Spikepingpong: High-frequency spike vision-based robot learning for precise striking in table tennis game,

    H. Wang, C. Hou, X. Li, Y . Fu, C. Li, N. Chen, G. Dai, J. Liu, T. Huang, and S. Zhang, “Spikepingpong: High-frequency spike vision-based robot learning for precise striking in table tennis game,” arXiv preprint arXiv:2506.06690, 2025

    arXiv 2025

  22. [22]

    Unified control framework for real- time interception and obstacle avoidance of fast-moving objects with diffusion variational autoencoder,

    A. Dastider, H. Fang, and M. Lin, “Unified control framework for real- time interception and obstacle avoidance of fast-moving objects with diffusion variational autoencoder,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 13 883–13 890

    2024

  23. [23]

    Real- time motion generation and data augmentation for grasping moving objects with dynamic speed and position changes,

    K. Yamamoto, H. Ito, H. Ichiwara, H. Mori, and T. Ogata, “Real- time motion generation and data augmentation for grasping moving objects with dynamic speed and position changes,” in2024 IEEE/SICE International Symposium on System Integration (SII). IEEE, 2024, pp. 390–397

    2024

  24. [24]

    Binoc- ular visual tracking and grasping of a moving object with a 3d trajectory predictor,

    J. Fuentes-Pacheco, J. Ruiz-Ascencio, and J. Rend ´on-Mancha, “Binoc- ular visual tracking and grasping of a moving object with a 3d trajectory predictor,”Journal of applied research and technology, vol. 7, no. 3, pp. 259–273, 2009

    2009

  25. [25]

    Catching robot: Predicting the trajectory of a rolling ball using transformer,

    N. Lee, Y . Oh, and J. Moon, “Catching robot: Predicting the trajectory of a rolling ball using transformer,”IEEE Access, 2024

    2024

  26. [26]

    Neural motion prediction for in-flight uneven object catching,

    H. Yu, D. Guo, H. Yin, A. Chen, K. Xu, Z. Chen, M. Wang, Q. Tan, Y . Wang, and R. Xiong, “Neural motion prediction for in-flight uneven object catching,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 4662– 4669

    2021

  27. [27]

    Impact-aware bimanual catching of large-momentum objects,

    L. Yan, T. Stouraitis, J. Moura, W. Xu, M. Gienger, and S. Vijayaku- mar, “Impact-aware bimanual catching of large-momentum objects,” IEEE Transactions on Robotics, vol. 40, pp. 2543–2563, 2024

    2024

  28. [28]

    Dipp: Discriminative impact point predictor for catching diverse in-flight objects,

    N. H. Nguyen, K. Shibata, and T. Matsubara, “Dipp: Discriminative impact point predictor for catching diverse in-flight objects,”arXiv preprint arXiv:2509.15254, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Agile catching with whole-body mpc and blackbox policy learning,

    S. Abeyruwan, A. Bewley, N. M. Boffi, K. M. Choromanski, D. B. D’Ambrosio, D. Jain, P. R. Sanketi, A. Shankar, V . Sindhwani, S. Singhet al., “Agile catching with whole-body mpc and blackbox policy learning,” inLearning for Dynamics and Control Conference. PMLR, 2023, pp. 851–863

    2023

  30. [30]

    Catching objects with a robot arm using model predictive control,

    T. Gold, R. R ¨omer, A. V ¨olz, and K. Graichen, “Catching objects with a robot arm using model predictive control,” in2022 American Control Conference (ACC). IEEE, 2022, pp. 1915–1920

    2022

  31. [31]

    Retro: Reactive trajectory optimiza- tion for real-time robot motion planning in dynamic environments,

    A. Dastider, H. Fang, and M. Lin, “Retro: Reactive trajectory optimiza- tion for real-time robot motion planning in dynamic environments,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 8764–8770

    2024

  32. [32]

    Catch the ball: Accurate high-speed motions for mobile manipulators via inverse dynamics learning,

    K. Dong, K. Pereida, F. Shkurti, and A. P. Schoellig, “Catch the ball: Accurate high-speed motions for mobile manipulators via inverse dynamics learning,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 6718–6725

    2020

  33. [33]

    Modular neural network policies for learning in-flight object catching with a robot hand-arm system,

    W. Hu, F. Acero, E. Triantafyllidis, Z. Liu, and Z. Li, “Modular neural network policies for learning in-flight object catching with a robot hand-arm system,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 944–951

    2023

  34. [34]

    Dnn task assignment in uav networks: A generative ai enhanced multi-agent reinforcement learning approach,

    X. Tang, Q. Chen, W. Weng, B. Liao, J. Wang, X. Cao, and X. Li, “Dnn task assignment in uav networks: A generative ai enhanced multi-agent reinforcement learning approach,”IEEE Internet of Things Journal, 2025

    2025

  35. [35]

    Learning multi-agent loco-manipulation for long-horizon quadrupedal pushing,

    Y . Feng, C. Hong, Y . Niu, S. Liu, Y . Yang, W. Yu, T. Zhang, J. Tan, and D. Zhao, “Learning multi-agent loco-manipulation for long-horizon quadrupedal pushing,”arXiv preprint arXiv:2411.07104, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning,

    M. Mittal, P. Roth, J. Tigue, A. Richard, and O. Z. et al., “Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning,”arXiv preprint arXiv:2511.04831, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  38. [38]

    skrl: Modular and flexible library for reinforcement learning,

    A. Serrano-Mu ˜noz, D. Chrysostomou, S. Bøgh, and N. Arana- Arexolaleiba, “skrl: Modular and flexible library for reinforcement learning,”Journal of Machine Learning Research, vol. 24, no. 254, pp. 1–9, 2023

    2023

  39. [39]

    Rectified linear units improve restricted boltzmann machines,

    V . Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines,” inProceedings of the 27th international con- ference on machine learning (ICML-10), 2010, pp. 807–814

    2010

  40. [40]

    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

    Pith/arXiv arXiv 2017