Pith. sign in

REVIEW 3 major objections 5 minor 34 references

RoboBallet claims that a single GNN policy trained with reinforcement learning on randomized scenes can jointly solve task allocation, scheduling, IK selection, and motion planning for up to eight 7-DoF arms and 40 reaching tasks, and can g

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

2026-08-05 05:33 UTC pith:ABEZNRUZ

load-bearing objection A genuinely scalable learned planner for multi-robot reaching, but the 'arbitrary obstacle geometries' claim is broader than the cuboid-only evidence supports. the 3 major comments →

arxiv 2509.05397 v1 pith:ABEZNRUZ submitted 2025-09-05 cs.RO cs.LG

RoboBallet: Planning for Multi-Robot Reaching with Graph Neural Networks and Reinforcement Learning

Matthew Lai , Keegan Go , Zhibin Li , Torsten Kroger , Stefan Schaal , Kelsey Allen , Jonathan Scholz This is my paper
classification cs.RO cs.LG
keywords multi-robot task and motion planninggraph neural networksreinforcement learningtask allocationschedulingcollision-free motion planningzero-shot generalizationworkcell layout optimization
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.

The paper tries to show that the expensive combinatorial problem of coordinating many robot arms in a cluttered shared workspace—deciding which robot does which task, in what order, with which inverse-kinematics solution, along which collision-free path—can be collapsed into a single learned control policy. The policy is a graph neural network trained by reinforcement learning on procedurally generated scenes, so planning time is just one forward pass rather than an online search. On the paper's largest test it coordinates eight 7-DoF arms on forty reaching tasks, producing plans in about 0.3 ms per step on an A100, and it transfers zero-shot to a hand-designed cell with robot placements, obstacle shapes, and task poses it has not seen. If true, this replaces a manual, hundreds-of-hours process with automated planning, and makes downstream uses like workcell layout optimization and rapid re-planning feasible.

Core claim

The central claim is that task allocation, task scheduling, IK solution selection, and motion planning do not need to be solved separately for dense multi-robot reaching: a single graph-structured policy, trained entirely in simulation on randomized workcells, can generate joint velocities for all robots simultaneously that solve all tasks. The paper demonstrates this up to eight 7-DoF robots and 40 tasks with no pre-assigned allocation or ordering. It reports that training-curve convergence is roughly invariant to task and robot count, and that trajectory durations are competitive with an RRT-Connect exhaustive-scheduling baseline at 8 IK samples per task, while the baseline plans each robo

What carries the argument

The carrying mechanism is a scene graph whose nodes are robots, tasks, and obstacle primitives, with bi-directional robot-robot edges for coordination and uni-directional task-to-robot and obstacle-to-robot edges for planning inputs. A graph neural network with learned edge, node, and global update functions processes this graph with weight sharing, keeping model size constant in scene size while runtime grows as N_robot^2 + N_robot*N_task + N_robot*N_obstacle. The policy is trained with a modified TD3 actor-critic algorithm plus Hindsight Experience Replay, using sparse task-completion rewards and collision penalties. Obstacle meshes are converted to cuboid primitives by approximate convex

Load-bearing premise

Every obstacle the policy meets must be representable as a modest number of cuboid primitives; the model never sees any other obstacle shape, so environments with thin or deeply concave obstacles that resist cuboid decomposition are outside what the trained policy can be trusted to plan around.

What would settle it

Take a workcell whose central obstacle is a thin curved sheet or deeply concave lattice that decomposes into hundreds of cuboid primitives, keep the robots and tasks within the training distribution, and measure zero-shot task-completion rate; the generalization claim would be overturned if the collision-free success rate drops sharply compared with cells whose obstacles decompose into a handful of cuboids.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A planner fast enough to run at 0.3 ms per planning step makes it practical to wrap in a black-box optimizer and tune robot placements for a fixed task set, yielding up to 33% faster execution in the paper's layout-optimization experiment.
  • Because the same policy can regenerate trajectories without retraining, a robot failure can be absorbed by replanning for the remaining arms, either ahead of time or on demand.
  • The reward function can be modified to add objectives such as returning to start or limiting acceleration without redesigning the planner, so the same framework can be adapted beyond pure reaching.
  • Scaling to more tasks or robots does not require more model parameters or exponentially more training steps; it only adds graph edges and roughly linear or quadratic compute.

Where Pith is reading between the lines

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

  • A natural next measurement is to test where performance degrades as obstacle primitives multiply; the paper's cuboid-decomposition assumption implies a testable cliff for thin or deeply concave obstacles that require hundreds of primitives.
  • If the learned interaction rules are truly modular across graph nodes, the same policy should transfer to heterogeneous robot teams after only modest morphology randomization during training, which the paper lists as future work.
  • The 2.5 cm and 15-degree task tolerance means reported trajectory costs exclude final precision docking; integrating a linear docking driver, as the paper notes, would add a small but nonzero cost that deployment studies should quantify.

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 RoboBallet, a GNN-based RL policy for joint task allocation, scheduling, IK solution selection, and motion planning for multiple 7-DoF arms in a shared workspace. The policy is trained in simulation on procedurally generated scenes with randomized robot placements, cuboid obstacles, and tasks sampled on obstacle surfaces. It outputs joint velocities at 10 Hz and is evaluated on up to 8 robots and 40 tasks, including a real four-robot workcell. The paper reports zero-shot generalization to a hand-designed aluminium-strut obstacle cell, inference at 0.3 ms/step on an A100, favorable training-scaling curves, a trajectory-cost comparison against RRT-Connect with exhaustive scheduling in a simplified single-allocation setting, and a workcell layout optimization use case.

Significance. If the main claims hold, this is a notable advance in multi-robot task and motion planning, showing that a single learned policy can amortize the joint planning problem and generalize to unseen configurations. The strengths include a principled graph representation with relative-pose edge features, a large-scale randomized training procedure, an open-source code release, and real-robot validation. The scalability results (Figs. 3 and 4) and the reported inference speed are compelling. However, the central zero-shot generalization claim to 'arbitrary obstacle geometries' is only tested on one hand-designed obstacle whose OBB decomposition is favorable, so the significance is partly conditional on additional evidence or a narrowing of the claim.

major comments (3)
  1. [Obstacle representation; Generalizability to hand-designed environments; Limitations and future work] The abstract and Introduction claim zero-shot generalization to 'arbitrary obstacle geometries,' but the model is trained only on exact cuboids and at evaluation obstacles are represented as oriented bounding boxes of V-HACD convex hulls. The only unseen-obstacle evaluation is a single aluminium-strut structure. No systematic test is provided across obstacle classes with varying decomposition error (e.g., thin tubes, concave shells, high OBB-to-mesh volume ratios). Because tasks are sampled on the true mesh surface while the obstacle node is the inflated OBB, the policy can receive task nodes lying inside the OBB volume, a situation never seen in training. The Limitations section concedes this representation 'may not be suitable' for many obstacles, directly qualifying the abstract's claim. This is load-bearing for the main generalization claim; either add a multi-shape evaluation or exp
  2. [Optimality of generated trajectories (Fig. 2)] The optimality comparison in Figure 2 uses tasks pre-assigned to one robot each and plans each robot with RRT-Connect in isolation, explicitly ignoring inter-robot collisions. The paper acknowledges this, but the concluding sentence 'competitive with the baseline approach at 8 IK samples per task' is then not supported as a statement about multi-robot trajectory cost. The baseline solves a strictly easier problem (fixed allocation, no inter-robot collision constraints). Please either provide a multi-robot-aware baseline (e.g., sequential planning with collision checking or prioritized planning) or reframe the result as a motion-cost sanity check on individually planned segments, not as an optimality benchmark for the joint multi-robot problem.
  3. [Modified TD3 (Materials and Methods)] The paper removes the delayed target network updates from TD3, stating that the stability contribution was minimal. This is contrary to the motivation in the original TD3 paper, where delayed updates are a key stabilizer. No ablation is given to support this deviation. Since the paper calls the algorithm 'modified TD3,' the deviation should be justified with an experiment or a reference to prior work, otherwise the reproducibility of the training procedure is weakened.
minor comments (5)
  1. [Efficient use of multiple robots (Fig. 4)] The text reports a 'reduction of 60%' when task completion time falls from approximately 7.5 s to 4.5 s. The actual reduction is (7.5 - 4.5)/7.5 = 40%. Please correct the percentage.
  2. [Table S5] The maximum number of training steps is written as '10×10 6' in the table; this should be 10^7 or 1e7. The formatting is likely a typesetting error but should be fixed.
  3. [Optimality of generated trajectories] The sentence 'RRT-Connect results are generated with each robot in isolation (other robots removed)' is important context, but it appears only in the Discussion of the baseline. Consider moving or repeating this caveat in the Results section where the comparison is introduced, so readers do not interpret Figure 2 as a full multi-robot comparison.
  4. [Introduction, para. 3] The phrase 'their runtime scales exponentially in the dimensionality of the configuration space and the complexity of the obstacles' is imprecise: RRT-style planners are not exponential in obstacle complexity in the same way as in dimension. Consider citing the specific complexity result or softening the wording.
  5. [Graph neural network core] In the text following Eq. (11), 'learned node embedding E_n and edge embedding E_e functions respectively' is confusing because both node and edge features are embedded. Clarify the sentence to avoid implying the edge features use the node embedding function.

Circularity Check

0 steps flagged

No significant circularity: zero-shot generalization is evaluated on held-out configurations and the derivation is self-contained.

full rationale

The paper's central claim is empirical: a GNN policy trained with TD3/HER on procedurally generated environments (randomized robot placements, cuboid obstacles, tasks) is evaluated on hand-designed evaluation cells with fixed robot placements, real-world obstacle meshes, and pre-sampled task sets explicitly not used in training (Fig. 3C-D, Movie 1). The predicted outputs are joint velocities integrated by a kinematic simulator with collision checks and task-completion tolerances; no evaluation metric is fitted back into the model. Hyperparameters (C_col, gamma, learning rates) were tuned with Vizier on training-like environments and are standard RL constants, not quantities calibrated to the reported evaluation outcomes. The obstacle representation pipeline (V-HACD decomposition to convex hulls, then oriented bounding boxes, Fig. 6) is a lossy approximation acknowledged in the Limitations section; this is a generalization/validity concern, not circularity, because training and evaluation share the same primitive-node representation and the claim about arbitrary geometries is an empirical extrapolation tested on one strut-cell obstacle. The trajectory-optimality comparison against RRT-Connect is explicitly qualified (tolerance 2.5 cm / 15°, single-robot baseline), and the differences are disclosed rather than concealed. There are no load-bearing self-citations, no imported uniqueness theorems, and no fitted quantity renamed as a prediction. The paper's contributions—task allocation, scheduling, IK selection, and motion planning via one learned policy—are not equivalent by construction to any of the paper's inputs; they are learned behaviors verified on held-out configurations. Hence no circular step is present.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on several domain assumptions: the kinematic simulator is a faithful stand-in for real robots; the cuboid decomposition captures obstacles well enough; the reward/HER design shapes minimal-time behavior; and the simplified RRT-Connect baseline is a meaningful proxy for trajectory optimality. No new physical entities are introduced. The free parameters (collision penalty, tolerance, dwell time, discount) were hand-chosen or tuned and materially affect the measured outcomes.

free parameters (4)
  • Collision penalty coefficient C_col = 15.0
    Tuned with the Vizier black-box optimizer; directly shapes the trade-off between task progress and collision avoidance, central to the collision-free claim.
  • Task completion tolerance (distance/angle) = 2.5 cm / 15 deg
    Hand-chosen; determines measured solution rates and trajectory costs, and differs from the exact pose used in the RRT-Connect baseline, complicating the optimality comparison.
  • Dwell time at each task = 0.5 s
    Hand-chosen; simulates work being done, affects task scheduling balance and the execution-time metric in layout optimization.
  • Discount factor gamma = 0.94
    Tuned with Vizier; sets temporal credit assignment for the RL objective and therefore affects the emergent behavior.
axioms (4)
  • domain assumption The kinematic simulator with velocity/acceleration clamping and collision-zeroing faithfully predicts real robot behavior
    Used in Methods Environment; the collision-free property of generated trajectories is enforced by zeroing colliding velocity commands, so real-world applicability depends on the simulator's fidelity.
  • domain assumption RRT-Connect with shortcutting gives near-optimal single-robot trajectories for the baseline comparison
    Used in Results 'Optimality of generated trajectories'; the baseline plans each robot in isolation and ignores inter-robot collisions, making the 'competitive' claim conditional on this simplification.
  • domain assumption A task with at least one collision-free IK solution is potentially solvable by the planner
    Stated in Methods 'Task sampling': path existence is not verified, so some tasks may be infeasible, which can depress task-solution-rate metrics independently of policy quality.
  • domain assumption The sparse reward (score difference + collision penalty) with HER shapes optimal execution-time behavior
    Used in Methods 'Reward' and 'Hindsight Experience Replay'; the paper chooses this reward design and does not prove it induces minimal-time trajectories.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 16562 in / 14184 out tokens · 135064 ms · 2026-08-05T05:33:05.498176+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of RoboBallet: Planning for Multi-Robot Reaching with Graph Neural Networks and Reinforcement Learning." pith.science (2026). https://pith.science/paper/ABEZNRUZ

@misc{pith2026250905397,
  author       = {Pith},
  title        = {Pith review of: RoboBallet: Planning for Multi-Robot Reaching with Graph Neural Networks and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ABEZNRUZ}},
  note         = {Machine review of arXiv:2509.05397}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Modern robotic manufacturing requires collision-free coordination of multiple robots to complete numerous tasks in shared, obstacle-rich workspaces. Although individual tasks may be simple in isolation, automated joint task allocation, scheduling, and motion planning under spatio-temporal constraints remain computationally intractable for classical methods at real-world scales. Existing multi-arm systems deployed in the industry rely on human intuition and experience to design feasible trajectories manually in a labor-intensive process. To address this challenge, we propose a reinforcement learning (RL) framework to achieve automated task and motion planning, tested in an obstacle-rich environment with eight robots performing 40 reaching tasks in a shared workspace, where any robot can perform any task in any order. Our approach builds on a graph neural network (GNN) policy trained via RL on procedurally-generated environments with diverse obstacle layouts, robot configurations, and task distributions. It employs a graph representation of scenes and a graph policy neural network trained through reinforcement learning to generate trajectories of multiple robots, jointly solving the sub-problems of task allocation, scheduling, and motion planning. Trained on large randomly generated task sets in simulation, our policy generalizes zero-shot to unseen settings with varying robot placements, obstacle geometries, and task poses. We further demonstrate that the high-speed capability of our solution enables its use in workcell layout optimization, improving solution times. The speed and scalability of our planner also open the door to new capabilities such as fault-tolerant planning and online perception-based re-planning, where rapid adaptation to dynamic task sets is required.

Figures

Figures reproduced from arXiv: 2509.05397 by Jonathan Scholz, Keegan Go, Kelsey Allen, Matthew Lai, Stefan Schaal, Torsten Kroger, Zhibin Li.

Figure 1
Figure 1. Figure 1: Reinforcement learning setup Environment The environment is implemented using a kinematic simulator. Each action applied ad￾vances the simulation by 0.1 seconds, and each action consists of desired joint velocities for all joints of all robots. We apply acceleration and velocity limiting based on robot kinematic limits. A task is considered solved if the end effector of a robot moves to within an Euclidean… view at source ↗
Figure 1
Figure 1. Figure 1: Training and evaluation pipeline. (A) The training pipeline. Each episode starts with an empty table, upon which we add random cuboid ob￾stacles, randomize robot placements, and randomly sample end effector tasks. The GNN actor interacts with the kinematic simulator to try to complete the tasks. The trajectories are stored in a replay buffer and used to train a critic and policy function using TD3. (B) The… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison against RRT-Connect + Exhaustive Task Scheduling Baseline in a simplified set￾ting. Each colour represents one task set with 20 tasks (5 per robot). All columns except ’RoboBallet’ uses RRT￾Connect and shortcutting for motion planning. The ’Random’ column averages over all possible task schedules. The ’Opt [N] IK’ columns uses N IK solutions per task, and the optimal task schedule and IK solutio… view at source ↗
Figure 3
Figure 3. Figure 3: Training Curves for Scalability and Generalizability. (A) Training curves for 10 versus 40 tasks, four robots. (B) Training curves for four versus eight robots, 40 tasks. Note that the number of training steps required to reach convergence does not scale exponentially. However, in terms of asymptotic time complexity, the amount of computation required is linear in the number of tasks, and quadratic in the … view at source ↗
Figure 4
Figure 4. Figure 4: Task completion time scaling with the number of robots (optimized placements). Shortest trajectory execution times achieved with the 10 evaluation configurations using optimized placements of four to eight robots. Each color represents one evaluation configuration [PITH_FULL_IMAGE:figures/full_fig_p042_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Obstacle mesh decomposition pipeline. (A) Starting from the CAD model of the obstacle. (B) Result of decomposition into convex polyhedrons using V-HACD. (C) Oriented bounding boxes computed from the convex polyhedrons using (22) [PITH_FULL_IMAGE:figures/full_fig_p044_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

34 extracted references · 30 canonical work pages · 1 internal anchor

  1. [1]

    Pellegrinelli, N

    S. Pellegrinelli, N. Pedrocchi, L. M. Tosatti, A. Fischer, T. Tolio, Multi-robot spot- welding cells for car-body assembly: Design and motion planning,Robotics and Computer-Integrated Manufacturing44, 97 (2017)

    work page 2017

  2. [2]

    Yamada, S

    Y . Yamada, S. Nagamatsu, Y . Sato, Development of multi-arm robots for automo- bile assembly,Proceedings of 1995 IEEE International Conference on Robotics and Automation(IEEE, 1995), vol. 3, pp. 2224–2229

    work page 1995

  3. [3]

    H. Chen, T. Fuhlbrigge, X. Li, Automated industrial robot path planning for spray painting process: a review,2008 IEEE International Conference on Automation Sci- ence and Engineering(IEEE, 2008), pp. 522–527

    work page 2008

  4. [4]

    V . N. Hartmann, A. Orthey, D. Driess, O. S. Oguz, M. Toussaint, Long-horizon multi-robot rearrangement planning for construction assembly,IEEE Transactions on Robotics39, 239 (2022)

    work page 2022

  5. [5]

    Z. Xian, P. Lertkultanon, Q.-C. Pham, Closed-chain manipulation of large objects by multi-arm robotic systems,IEEE Robotics and Automation Letters2, 1832 (2017)

    work page 2017

  6. [6]

    LaValle, Rapidly-exploring random trees: A new tool for path planning,Research Report 9811(1998)

    S. LaValle, Rapidly-exploring random trees: A new tool for path planning,Research Report 9811(1998)

    work page 1998

  7. [7]

    S. M. LaValle, J. J. Kuffner, Rapidly-exploring random trees: Progress and prospects: Steven m. lavalle, iowa state university, a james j. kuffner, jr., university of tokyo, tokyo, japan,Algorithmic and computational roboticspp. 303–307 (2001)

    work page 2001

  8. [8]

    Karaman, E

    S. Karaman, E. Frazzoli, Sampling-based algorithms for optimal motion planning, The international journal of robotics research30, 846 (2011)

    work page 2011

  9. [9]

    Canny,The complexity of robot motion planning(MIT press, 1988)

    J. Canny,The complexity of robot motion planning(MIT press, 1988). 32

    work page 1988

  10. [10]

    Vega-Brown, N

    W. Vega-Brown, N. Roy, Task and motion planning is pspace-complete,Proceedings of the AAAI Conference on Artificial Intelligence(2020), vol. 34, pp. 10385–10392

    work page 2020

  11. [11]

    G. B. Mathews, On the partition of numbers,Proceedings of the London Mathemati- cal Society1, 486 (1896)

    work page

  12. [12]

    H. Ha, J. Xu, S. Song, Learning a decentralized multi-arm motion planner,arXiv preprint arXiv:2011.02608(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  13. [13]

    T. Pan, A. M. Wells, R. Shome, L. E. Kavraki, A general task and motion planning framework for multiple manipulators,2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE, 2021), pp. 3168–3174

    work page 2021

  14. [14]

    LeCun,et al., Handwritten digit recognition with a back-propagation network, Advances in neural information processing systems2(1989)

    Y . LeCun,et al., Handwritten digit recognition with a back-propagation network, Advances in neural information processing systems2(1989)

    work page 1989

  15. [15]

    J. J. Kuffner, S. M. LaValle, Rrt-connect: An efficient approach to single-query path planning,Proceedings 2000 ICRA. Millennium Conference. IEEE International Con- ference on Robotics and Automation. Symposia Proceedings (Cat. No. 00CH37065) (IEEE, 2000), vol. 2, pp. 995–1001

    work page 2000

  16. [16]

    Golovin,et al., eds.,Google Vizier: A Service for Black-Box Optimization(2017)

    D. Golovin,et al., eds.,Google Vizier: A Service for Black-Box Optimization(2017)

    work page 2017

  17. [17]

    Song,et al., The vizier gaussian process bandit algorithm,arXiv preprint arXiv:2408.11527(2024)

    X. Song,et al., The vizier gaussian process bandit algorithm,arXiv preprint arXiv:2408.11527(2024)

    Pith/arXiv arXiv 2024

  18. [18]

    Velickovic,et al., Graph attention networks,stat1050, 10 (2017)

    P. Velickovic,et al., Graph attention networks,stat1050, 10 (2017)

    work page 2017

  19. [19]

    P. W. Battaglia,et al., Relational inductive biases, deep learning, and graph networks, arXiv preprint arXiv:1806.01261(2018). 33

    Pith/arXiv arXiv 2018

  20. [20]

    Fujimoto, H

    S. Fujimoto, H. Hoof, D. Meger, Addressing function approximation error in actor- critic methods,International conference on machine learning(PMLR, 2018), pp. 1587–1596

    work page 2018

  21. [21]

    e. a. John W. Ratcliff, Khaled Mamou, V oxelized hierarchical convex decomposition - v-hacd version 4 (2023)

    work page 2023

  22. [22]

    Chang, B

    C.-T. Chang, B. Gorissen, S. Melchior, Fast oriented bounding box optimization on the rotation group so (3, ),ACM Transactions on Graphics (TOG)30, 1 (2011)

    work page 2011

  23. [23]

    D. M. Olsson, L. S. Nelson, The nelder-mead simplex procedure for function mini- mization,Technometrics17, 45 (1975)

    work page 1975

  24. [24]

    Y . Zhou, C. Barnes, J. Lu, J. Yang, H. Li, On the continuity of rotation representations in neural networks,Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2019), pp. 5745–5753

    work page 2019

  25. [25]

    Andrychowicz,et al., Hindsight experience replay,Advances in neural informa- tion processing systems30(2017)

    M. Andrychowicz,et al., Hindsight experience replay,Advances in neural informa- tion processing systems30(2017)

    work page 2017

  26. [26]

    Silver,et al., Deterministic policy gradient algorithms,International conference on machine learning(Pmlr, 2014), pp

    D. Silver,et al., Deterministic policy gradient algorithms,International conference on machine learning(Pmlr, 2014), pp. 387–395

    work page 2014

  27. [27]

    B. T. Polyak, A new method of stochastic approximation type,Avtomatika i tele- mekhanikapp. 98–107 (1990)

    work page 1990

  28. [28]

    J. L. Ba, J. R. Kiros, G. E. Hinton, Layer normalization,arXiv preprint arXiv:1607.06450(2016)

    Pith/arXiv arXiv 2016

  29. [29]

    Hendrycks, K

    D. Hendrycks, K. Gimpel, Gaussian error linear units (gelus),arXiv preprint arXiv:1606.08415(2016). 34

    Pith/arXiv arXiv 2016

  30. [30]

    Bradbury,et al., JAX: composable transformations of Python+NumPy programs (2018)

    J. Bradbury,et al., JAX: composable transformations of Python+NumPy programs (2018)

    work page 2018

  31. [31]

    Godwin*,et al., Jraph: A library for graph neural networks in jax

    J. Godwin*,et al., Jraph: A library for graph neural networks in jax. (2020)

    work page 2020

  32. [32]

    Heek,et al., Flax: A neural network library and ecosystem for JAX (2023)

    J. Heek,et al., Flax: A neural network library and ecosystem for JAX (2023)

    work page 2023

  33. [33]

    Cassirer,et al., Reverb: A framework for experience replay (2021)

    A. Cassirer,et al., Reverb: A framework for experience replay (2021)

    work page 2021

  34. [34]

    Todorov, T

    E. Todorov, T. Erez, Y . Tassa, Mujoco: A physics engine for model-based control, 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems(IEEE, 2012), pp. 5026–5033. ACKNOWLEDGEMENTS Funding:This work was funded by Google DeepMind and Intrinsic. Author contributions:ML, KA conceptualized, designed, and conducted the experi- ments. ML, KA,...

    work page 2012