Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

Neural models can replace the physics core of a robot simulator and stay accurate over thousands of steps.

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 ↗

Neural dynamics models trained on random simulator data act as a general robot physics backend, supporting policy learning and real-world fine-tuning with stable long-horizon prediction.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A strong, well-engineered neural simulator paper; the torque-trajectory gap is real but testable, and the empirical package justifies serious review. the 3 major comments →

arxiv 2508.15755 v1 pith:YJR77OQL submitted 2025-08-21 cs.RO cs.AIcs.GRcs.LG

Neural Robot Dynamics

Jie Xu , Eric Heiden , Iretiayo Akinola , Dieter Fox , Miles Macklin , Yashraj Narang This is my paper
classification cs.RO cs.AIcs.GRcs.LG
keywords neural robot dynamicsgeneralizable simulationarticulated rigid bodiescontact constraintsrobot-centric state representationlong-horizon stabilitysim-to-real transfer
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 reading

This paper proposes NeRD, a learned, robot-specific dynamics model that replaces only the low-level forward-dynamics and contact solvers of an analytical simulator rather than the whole engine. The model takes, as input, the same application-agnostic intermediate quantities a classical solver uses, expressed in the robot's own base frame: recent robot states, contact measurements at fixed points, joint torques, and the gravity direction. This representation is what the authors claim gives the model its generalization across tasks, environments, and controllers, plus long-horizon stability. If correct, it would mean robots could carry a pretrained neural digital twin that is stable for thousands of steps, supports policy learning entirely inside the neural engine, and can be fine-tuned from a small amount of real-world data.

Core claim

The paper claims that a learned model can act as an interchangeable backend solver inside a classical articulated-rigid-body simulator. NeRD consumes the robot's state, contact events at pre-specified points, joint torques, and the base-frame gravity vector over a ten-step history, and outputs the relative state change to the next macro-step. Because the inputs are the same intermediate quantities used by classical dynamics and contact solvers, the model does not need to memorize a specific scene or controller. Because the representation is robot-centric, it is invariant under translation and rotation about the gravity axis, which the authors argue is what lets a model trained on random traj

What carries the argument

The central mechanism is the hybrid prediction framework: collision detection and the low-level controller remain in the analytical simulator, while NeRD replaces the solver that integrates dynamics under contact. Its state representation is robot-centric, expressing all inputs and the predicted state difference in the moving base frame and treating gravity as an external vector in that frame, which enforces spatial invariance and shrinks the state space. A causal Transformer over the ten-step history maps these quantities to the next state difference; the relative prediction target keeps outputs small, and the history provides smoothed velocity information that stabilizes training and rollo

Load-bearing premise

A fixed set of contact points plus the torque at the start of a macro-step fully determines what happens to the robot over that step, so every interaction with the environment must be visible through those measurements.

What would settle it

Take a NeRD model trained with a fixed set of contact points on a robot link, then place an obstacle that touches that link at a different, unprescribed point; if the neural rollout diverges from the analytical simulator within a few hundred steps, the contact representation does not capture all relevant interaction.

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

If this is right

  • One pretrained NeRD model can support policy learning for tasks whose state distributions were never seen in its random training data, and the learned policies achieve rewards essentially matching those of a ground-truth analytical simulator.
  • Because NeRD predicts the macro-step state difference directly, it can bypass finer substepping at test time, giving a speed advantage while remaining stable over long horizons.
  • The same trained model works with different low-level controllers, such as joint-torque and joint-position control, without retraining.
  • A NeRD model pretrained on simulation can be fine-tuned from a small real-world dataset in minutes, outperforming the analytical simulator on that data and matching specialized single-body models.
  • Fine-tuning makes a continual digital-twin workflow plausible: each robot's simulator can be updated as the robot wears or the environment changes.

Where Pith is reading between the lines

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

  • The robot-centric invariance idea suggests that other symmetry structures of robot dynamics, such as reflection or repetitive joint topology, could be encoded the same way to further reduce data requirements.
  • The fixed-contact-point assumption implies the model's generalization is bounded by the chosen contact layout; adding learned or adaptive contact-point selection could extend NeRD to arbitrary impact locations.
  • Direct macro-step prediction may also serve as a differentiable world model for policy optimization through gradient-based backpropagation, beyond the reinforcement-learning use shown in the paper.
  • The fact that sequence modeling, rather than the stateless physics solver, improves accuracy hints that learned simulators may benefit from explicitly modeling unobserved substep behavior rather than only fitting state transitions.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. NeRD is a learned robot-specific dynamics model that replaces the low-level forward-dynamics and contact solvers of an analytical simulator while retaining the simulator's collision detection and low-level controller modules. Its input is a robot-centric, body-frame state representation built from robot state, contact events, joint torques, and gravity over a history window of h=10; a causal Transformer predicts the state difference over one macro-step. The model is pretrained on random trajectories generated by Warp and then evaluated on long-horizon prediction (Cartpole, Ant), contact generalization (Double Pendulum), RL policy learning for four robots, zero-shot sim-to-real Franka reach, and fine-tuning on real cube-tossing data. The paper reports low errors, close reward parity between NeRD and the analytical simulator in most tasks, and real-world transfer, with ablations supporting the main design choices.

Significance. If the central claims hold, NeRD is a significant step toward generalizable neural simulators for articulated rigid-body robots: it is tested on six systems, uses a modular hybrid framework, shows long-horizon stability, and demonstrates that policies trained exclusively inside the learned simulator can transfer to an analytical simulator and to a real Franka. The paper also provides concrete quantitative evidence: 2048-trajectory evaluations, ablations, and real-world fine-tuning with comparison to specialized baselines. The main caveats are the underspecified interaction between macro-step prediction and PD controllers, and one task with a large reward discrepancy. These do not invalidate the overall approach but do require targeted revision before the generalization claims are fully supported.

major comments (3)
  1. [§3.2, Fig. 2, Appendix A, Appendix C.1.3] The model input contains the joint torque τ_t only at the macro-step boundary, but the model may span multiple analytical substeps (Appendix A) and the Franka/ANYmal policies are executed through joint-position PD controllers (Appendix C.1.3). If the PD controller is evaluated at the physics substep rate, the torque trajectory inside the macro-step is not determined by the model's input: the same input vector corresponds to many possible next states. Training data were generated with random torque sequences (§4), not with PD feedback, so the history of boundary torques does not encode the intra-step torque law. This is a load-bearing gap for the claimed generalization to customized low-level controllers. Please report the control update rate in all position-controlled experiments; if it is faster than the macro-step, either include the torque trajectory/target positions and gains in the
  2. [§5.3, Table 1] The text states that NeRD-trained policies achieve 'remarkably similar rewards' when deployed in the NeRD and ground-truth simulators. However, the Ant Spin Tracking row reports a +17.21% reward difference (GT 2624.7 ± 641.0 vs. NeRD 3076.5 ± 433.5), the largest discrepancy of the seven tasks and in the optimistic direction. This is relevant to the central claim of zero-shot deployment in the analytical simulator. The paper should either explain this outlier or qualify the claim; as written, the 'remarkably similar' statement is overstated for this task.
  3. [§5.1, §5.3] The abstract claims NeRD is 'stable and accurate over a thousand simulation steps.' In §5.1 the 1000-step evidence is Cartpole only, which is contact-free; the contact-rich Ant evaluation is 500 steps. The ANYmal tasks run 1000 episodes but report reward, not per-step state error. Please provide a per-step state-error evaluation for at least one contact-rich system at 1000 steps, or soften the claim to reflect the systems actually evaluated at that horizon.
minor comments (4)
  1. [§5.2, Table 3] The Double Pendulum contact-generalization results are reported without comparison to an analytical baseline or an ablation that lacks contact information. Since the claim is about the necessity of contact encoding, adding such a baseline would strengthen the quantitative support.
  2. [Appendix C.6] The computation-speed comparison (Warp 28K FPS vs. NeRD 46K FPS for Ant) is presented with appropriate caveats. It would be useful to report the same comparison for the position-controlled Franka/ANYmal systems, since PD evaluation rate interacts with the macro-step issue raised above.
  3. [§4] The training-data section states that random trajectories are generated 'within the robot's motor torque limits,' but no details are given on how torque sequences are temporally correlated (e.g., white noise vs. smooth random walks) or whether they are held constant over substeps. This matters for understanding what the model can infer from τ_t in the history window.
  4. [Appendix C.1.3] The phrase 'converted from the target joint positions via the joint-position PD controller' is important but underspecified. Please state the PD gains, the control frequency, and whether the controller is evaluated at the macro-step rate or the physics-substep rate.

Circularity Check

0 steps flagged

No significant circularity: NeRD is trained on and evaluated against an external analytical simulator and real-world data.

full rationale

The central derivation is not circular. NeRD is a learned surrogate for the low-level dynamics and contact solver of an analytical simulator. Training data are generated by the Warp/Featherstone simulator (§4: "We generate the training datasets for NeRD in a task-agnostic manner using Warp with the Featherstone solver"), and the learned model is then tested on held-out trajectories of that same analytical simulator, on RL tasks never seen during training, and on real-world cube-tossing data (§5). This is a standard supervised-learning evaluation against an external benchmark, not a self-referential prediction: the target state differences are produced by an independent numerical solver, not by the model's own inputs or by a fitted parameter renamed as a prediction. The self-citations to Warp [24] and NeuralSim [42] are not load-bearing: Warp is an open-source analytical simulator whose outputs are generated independently of the trained model, and NeuralSim is discussed only as related work. The robot-centric state representation and relative-state prediction are architectural choices that improve learning, not circular definitions of the target. The only substantive caveat is the macro-step determinism assumption in Appendix A ("Multi-Substep Prediction"): if a PD controller updates inside the macro-step, the model input may not uniquely determine the next state. That is a correctness/generalization limitation, not circularity, because the model is not fitted to a quantity that is also an input. Hence the paper's central claims have independent empirical content.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities; NeRD is a learned parametric function. The free parameters listed are the hand-set values that most directly affect the learned representation and the pretraining data. The axioms are the domain assumptions about the sufficiency of the intermediate-quantity representation, the completeness of the point-contact model, the spatial symmetry exploited by the robot-centric frame, the coverage of random training data, and the compressibility of substep dynamics into a macro-step.

free parameters (4)
  • history window size h = 10
    Chosen by hand; ablation (Fig. 7c) shows h=10 generally best, and h=20 occasionally causes exploded training loss.
  • contact distance threshold xi = max(4 * contact thickness, 0.1)
    Fixed across all experiments; paper states it is flexible and not task-tuned (Appendix A).
  • Warp contact and inertia parameters for cube-tossing pretraining = manually tuned
    Used to replicate observed real-world cube dynamics in simulation before fine-tuning (Appendix C.4).
  • Transformer hyperparameters (embedding 192/384, 6 layers, 12 heads, learning rate schedule) = see Table 2
    Fixed across robots except embedding size; chosen by hand during development.
axioms (5)
  • domain assumption The next macro-step state is a function of the current state, contact info, joint torques, gravity, and the previous h steps; no additional environment or controller state is needed.
    Introduced in Section 3.2 where the model maps {s_k, C_k, tau_k} to Delta s_{t+1}.
  • domain assumption A fixed set of pre-specified contact points p_i^0 on the robot, with the closest non-robot point, normal, and distance, completely characterizes contact interaction with any environment.
    Appendix A defines C_t this way and the paper relies on it for generalization across environments; untested for dense or unpredictable multi-contact manifolds.
  • domain assumption Rigid-body dynamics with a uniform gravity field are invariant under translation and rotation about the gravity axis, and this symmetry is losslessly encoded by expressing inputs and outputs in the robot base frame plus the gravity vector.
    Section 3.3 uses this to justify the robot-centric state representation.
  • domain assumption Random task-agnostic trajectories provide sufficient coverage of the state-action distribution for downstream tasks.
    Stated in Section 4 and acknowledged as potentially failing for high-DoF robots in Section 6 (Limitations).
  • domain assumption The torque command at the start of a macro-step determines the torques applied over the entire step (or the model can recover this from history); the learned model need not model substeps.
    Multi-Substep Prediction in Appendix A explains NeRD directly predicts the macro-step difference, but the input only includes tau_t, not the substep torque trajectory under feedback controllers.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Robot Dynamics." pith.science (2026). https://pith.science/paper/YJR77OQL

@misc{pith2026250815755,
  author       = {Pith},
  title        = {Pith review of: Neural Robot Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJR77OQL}},
  note         = {Machine review of arXiv:2508.15755}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate and efficient simulation of modern robots remains challenging due to their high degrees of freedom and intricate mechanisms. Neural simulators have emerged as a promising alternative to traditional analytical simulators, capable of efficiently predicting complex dynamics and adapting to real-world data; however, existing neural simulators typically require application-specific training and fail to generalize to novel tasks and/or environments, primarily due to inadequate representations of the global state. In this work, we address the problem of learning generalizable neural simulators for robots that are structured as articulated rigid bodies. We propose NeRD (Neural Robot Dynamics), learned robot-specific dynamics models for predicting future states for articulated rigid bodies under contact constraints. NeRD uniquely replaces the low-level dynamics and contact solvers in an analytical simulator and employs a robot-centric and spatially-invariant simulation state representation. We integrate the learned NeRD models as an interchangeable backend solver within a state-of-the-art robotics simulator. We conduct extensive experiments to show that the NeRD simulators are stable and accurate over a thousand simulation steps; generalize across tasks and environment configurations; enable policy learning exclusively in a neural engine; and, unlike most classical simulators, can be fine-tuned from real-world data to bridge the gap between simulation and reality.

Figures

Figures reproduced from arXiv: 2508.15755 by Dieter Fox, Eric Heiden, Iretiayo Akinola, Jie Xu, Miles Macklin, Yashraj Narang.

Figure 1
Figure 1. Figure 1: We propose NeRD, learned robot-specific dynamics models for generalizable articulated rigid body simulation. We demonstrate our approach by training NeRD models on six diverse robotic systems, from left: Cartpole, Double Pendulum, Ant, Franka, ANYmal, Cube Toss. In this work, we address the problem of learning generalizable neural simulators for articulated rigid-body robots. We envision a future where eac… view at source ↗
Figure 2
Figure 2. Figure 2: Framework overview for Neural Robot Dynamics (NeRD). (a) Workflow of a classical robotics simulator. The quantities shaded in green are application-agnostic. (b) Hybrid prediction framework of the NeRD-integrated simulator. Inputs to NeRD are the robot-centric state representa￾tions (illustrated in (c)) within a history window. employs a hybrid prediction framework that replaces only the core physics compo… view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation of NeRD on long-horizon passive motions. Left: Full report of the measured errors. Right: 1000-step cartpole state trajectories simulated by NeRD and ground-truth simulator. Can NeRD reliably and accurately simulate long-horizon robotic trajectories (§5.1)? Does NeRD’s hybrid prediction framework enable it to generalize across different contact configurations (§5.2)? Can a single NeRD model gene… view at source ↗
Figure 4
Figure 4. Figure 4: Zero-shot sim-to-real transfer of a Franka reach policy. The real-world setup is shown in the left figure. The plot on the middle visualizes the evolution of distance-to-goal measurements when 10 NeRD-trained policies are executed, with a zoomed-in plot in the right. (a) Warp Simulator (b) Fine-tuned NeRD Simulator (c) Fine-tune v.s. Train from scratch [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Fine-tuning of a pretrained NeRD model on real-world cube-tossing data. (a-b) Cube-tossing trajectories simulated by the Warp simulator and by the fine-tuned NeRD simulator. The light-green frames are ground-truth poses. (c) Comparison of fine-tuning a pretrained NeRD model (red) against training a NeRD model from scratch (blue) on the real dataset. 5.5 Fine-tunability on Real-World Data: Cube Tossing We e… view at source ↗
Figure 6
Figure 6. Figure 6: Seven Double Pendulum contact configurations used for testing NeRD’s generaliz￾ability across different contact environments. C.3 Franka Reach Policy Learning with Joint-Torque Control In §5.3, we conduct RL policy-learning experiments to show the NeRD model’s generalizability to low-level controllers, where we apply different low-level controllers on different robots: joint-torque 17 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation Study. We evaluate ablation variants on two test cases: contact-free passive motion of the Double Pendulum and policy evaluation on the Ant running task. We normalize the errors by the error of NeRD (h = 10). (a) Ablations of different neural network architectures; (b) Ablations of other critical design decisions in NeRD. (c) Ablations on the history window size h. we conduct our study using two e… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation

    cs.RO 2026-03 unverdicted novelty 6.0

    SimDist pretrains world models in simulation and adapts them to real-world robots by updating only the latent dynamics model, enabling rapid improvement on contact-rich tasks where prior methods fail.

  2. Pretraining in Actor-Critic Reinforcement Learning for Locomotion

    cs.RO 2025-10 conditional novelty 6.0

    Pretraining an inverse dynamics model on task-agnostic exploration data and using its weights to initialize PPO actor and critic improves sample efficiency by ~37% and final performance by ~7% across nine simulated lo...

  3. CableRobotGraphSim: A Graph Neural Network for Modeling Partially Observable Cable-Driven Robot Dynamics

    cs.RO 2026-02 conditional novelty 5.0

    A fully learnable GNN predicts cable-driven tensegrity dynamics from partial observations and serves as the transition model for closed-loop MPPI navigation, beating differentiable physics baselines.

Reference graph

Works this paper leans on

52 extracted references · 47 canonical work pages · cited by 3 Pith papers · 1 internal anchor

  1. [1]

    O. M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, et al. Learning dexterous in-hand manipulation. The International Journal of Robotics Research , 2020

    work page 2020

  2. [2]

    T. Chen, M. Tippur, S. Wu, V . Kumar, E. Adelson, and P. Agrawal. Visual dexterity: In-hand reorientation of novel and complex object shapes. Science Robotics, 2023

    work page 2023

  3. [3]

    Haarnoja, B

    T. Haarnoja, B. Moran, G. Lever, S. H. Huang, D. Tirumala, J. Humplik, M. Wulfmeier, S. Tun- yasuvunakool, N. Y . Siegel, R. Hafner, M. Bloesch, K. Hartikainen, A. Byravan, L. Hasen- clever, Y . Tassa, F. Sadeghi, N. Batchelor, F. Casarini, S. Saliceti, C. Game, N. Sreendra, K. Patel, M. Gwira, A. Huber, N. Hurley, F. Nori, R. Hadsell, and N. Heess. Learn...

    work page 2024

  4. [4]

    Handa, A

    A. Handa, A. Allshire, V . Makoviychuk, A. Petrenko, R. Singh, J. Liu, D. Makoviichuk, K. Van Wyk, A. Zhurkevich, B. Sundaralingam, et al. Dextreme: Transfer of agile in-hand manipulation from simulation to reality. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023

    work page 2023

  5. [5]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, and J. Malik. Rma: Rapid motor adaptation for legged robots. 2021

    work page 2021

  6. [6]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomo- tion over challenging terrain. Science robotics, 2020

    work page 2020

  7. [7]

    B. Tang, I. Akinola, J. Xu, B. Wen, A. Handa, K. Van Wyk, D. Fox, G. S. Sukhatme, F. Ramos, and Y . Narang. Automate: Specialist and generalist assembly policies over diverse geometries. In Robotics: Science and Systems , 2024. 9

    work page 2024

  8. [8]

    N. Funk, C. Schaff, R. Madan, T. Yoneda, J. U. De Jesus, J. Watson, E. K. Gordon, F. Widmaier, S. Bauer, S. S. Srinivasa, et al. Benchmarking structured policies and policy optimization for real-world dexterous object manipulation. IEEE Robotics and Automation Letters , 2021

    work page 2021

  9. [9]

    J. Gu, F. Xiang, X. Li, Z. Ling, X. Liu, T. Mu, Y . Tang, S. Tao, X. Wei, Y . Yao, X. Yuan, P. Xie, Z. Huang, R. Chen, and H. Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In The Eleventh International Conference on Learning Representations , 2023

    work page 2023

  10. [10]

    X. Li, K. Hsu, J. Gu, K. Pertsch, O. Mees, H. R. Walke, C. Fu, I. Lunawat, I. Sieh, S. Kir- mani, S. Levine, J. Wu, C. Finn, H. Su, Q. Vuong, and T. Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Z. Liu, W. Liu, Y . Qin, F. Xiang, M. Gou, S. Xin, M. A. Roa, B. Calli, H. Su, Y . Sun, et al. Ocrtoc: A cloud-based competition and benchmark for robotic grasping and manipulation. IEEE Robotics and Automation Letters , 2021

    work page 2021

  12. [12]

    T. Du, A. Schulz, B. Zhu, B. Bickel, and W. Matusik. Computational multicopter design.ACM Trans. Graph., 2016

    work page 2016

  13. [13]

    M. Li, R. Antonova, D. Sadigh, and J. Bohg. Learning tool morphology for contact-rich manipulation tasks with differentiable simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , 2023

    work page 2023

  14. [14]

    J. Xu, T. Chen, L. Zlokapa, M. Foshey, W. Matusik, S. Sueda, and P. Agrawal. An End-to- End Differentiable Framework for Contact-Aware Robot Design. In Proceedings of Robotics: Science and Systems, 2021

    work page 2021

  15. [15]

    Pfrommer, M

    S. Pfrommer, M. Halm, and M. Posa. Contactnets: Learning discontinuous contact dynam- ics with smooth, implicit representations. In Proceedings of the 2020 Conference on Robot Learning, 2021

    work page 2020

  16. [16]

    Hoffman, J

    B. Hoffman, J. Cheng, C. Li, and S. Coros. Learning more with less: Sample efficient dynamics learning and model-based rl for loco-manipulation. arXiv preprint arXiv:2501.10499, 2025

    work page arXiv 2025

  17. [17]

    A. Ajay, J. Wu, N. Fazeli, M. Bauza, L. P. Kaelbling, J. B. Tenenbaum, and A. Rodriguez. Augmenting physical simulators with stochastic neural networks: Case study of planar pushing and bouncing. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018

    work page 2018

  18. [18]

    A. Zeng, S. Song, J. Lee, A. Rodriguez, and T. Funkhouser. Tossingbot: Learning to throw arbitrary objects with residual physics. IEEE Transactions on Robotics, 2020

    work page 2020

  19. [19]

    Janner, J

    M. Janner, J. Fu, M. Zhang, and S. Levine. When to trust your model: Model-based policy optimization. In Advances in Neural Information Processing Systems , 2019

    work page 2019

  20. [20]

    C. Li, A. Krause, and M. Hutter. Robotic world model: A neural network simulator for robust policy optimization in robotics, 2025

    work page 2025

  21. [21]

    Andriluka, B

    M. Andriluka, B. Tabanpour, C. D. Freeman, and C. Sminchisescu. Learned neural physics simulation for articulated 3d human pose reconstruction. In Computer Vision – ECCV 2024: 18th European Conference, Proceedings, Part LXXXIV, 2024

    work page 2024

  22. [22]

    Fussell, K

    L. Fussell, K. Bergamin, and D. Holden. Supertrack: motion tracking for physically simulated characters using supervised learning. ACM Trans. Graph., 40(6), 2021

    work page 2021

  23. [23]

    Hansen, X

    N. Hansen, X. Wang, and H. Su. Temporal difference learning for model predictive control. In International Conference on Machine Learning, PMLR , 2022. 10

    work page 2022

  24. [24]

    M. Macklin. Warp: A high-performance python framework for gpu simulation and graphics. https://github.com/nvidia/warp, 2022

    work page 2022

  25. [25]

    Bertiche, M

    H. Bertiche, M. Madadi, and S. Escalera. Neural cloth simulation. ACM Trans. Graph., 2022

    work page 2022

  26. [26]

    Y . Jin, D. Omens, Z. Geng, J. Teran, A. Kumar, K. Tashiro, and R. Fedkiw. A neural-network- based approach for loose-fitting clothing. arXiv preprint arXiv:2404.16896, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    Pfaff, M

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, and P. W. Battaglia. Learning mesh-based sim- ulation with graph networks. In International Conference on Learning Representations, 2021

    work page 2021

  28. [28]

    Ladick `y, S

    L. Ladick `y, S. Jeong, B. Solenthaler, M. Pollefeys, and M. Gross. Data-driven fluid simulations using regression forests. ACM Transactions on Graphics (TOG), 2015

    work page 2015

  29. [29]

    Sanchez-Gonzalez, J

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, and P. W. Battaglia. Learning to simulate complex physics with graph networks. In International Conference on Machine Learning, 2020

    work page 2020

  30. [30]

    J. Li, Y . Gao, J. Dai, S. Li, A. Hao, and H. Qin. Mpmnet: A data-driven mpm framework for dynamic fluid-solid interaction. IEEE Transactions on Visualization and Computer Graphics , 2024

    work page 2024

  31. [31]

    Li, Y .-L

    X. Li, Y .-L. Qiao, P. Y . Chen, K. M. Jatavallabhula, M. Lin, C. Jiang, and C. Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identifica- tion. International Conference on Learning Representations (ICLR) , 2023

    work page 2023

  32. [32]

    Jiang, J

    Y . Jiang, J. Sun, and C. K. Liu. Data-augmented contact model for rigid body simulation. In Proceedings of The 4th Annual Learning for Dynamics and Control Conference , 2022

    work page 2022

  33. [33]

    K. R. Allen, T. L. Guevara, Y . Rubanova, K. Stachenfeld, A. Sanchez-Gonzalez, P. Battaglia, and T. Pfaff. Graph network simulators can learn discontinuous, rigid contact dynamics. In Proceedings of The 6th Conference on Robot Learning , 2023

    work page 2023

  34. [34]

    K. R. Allen, Y . Rubanova, T. Lopez-Guevara, W. F. Whitney, A. Sanchez-Gonzalez, P. Battaglia, and T. Pfaff. Learning rigid dynamics with face interaction graph networks. In The Eleventh International Conference on Learning Representations , 2023

    work page 2023

  35. [35]

    Hafner, T

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi. Dream to control: Learning behaviors by latent imagination. In International Conference on Learning Representations , 2020

    work page 2020

  36. [36]

    Hafner, J

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Hansen, H

    N. Hansen, H. Su, and X. Wang. Td-mpc2: Scalable, robust world models for continuous control, 2024

    2024

  38. [38]

    T. M. Moerland, J. Broekens, A. Plaat, C. M. Jonker, et al. Model-based reinforcement learn- ing: A survey. F oundations and Trends® in Machine Learning, 2023

    work page 2023

  39. [39]

    Sanchez-Gonzalez, N

    A. Sanchez-Gonzalez, N. Heess, J. T. Springenberg, J. Merel, M. Riedmiller, R. Hadsell, and P. Battaglia. Graph networks as learnable physics engines for inference and control. In Inter- national conference on machine learning , 2018

    work page 2018

  40. [40]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. Karniadakis. Physics-informed neural networks: A deep learn- ing framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 2019

    work page 2019

  41. [41]

    Cuomo, V

    S. Cuomo, V . S. Di Cola, F. Giampaolo, G. Rozza, M. Raissi, and F. Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. 2022. 11

    work page 2022

  42. [42]

    Heiden, D

    E. Heiden, D. Millard, E. Coumans, Y . Sheng, and G. S. Sukhatme. Neuralsim: Augmenting differentiable simulators with neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021

    work page 2021

  43. [43]

    M. Kim, J. Yoon, D. Son, and D. Lee. Data-driven contact clustering for robot simulation. In 2019 International Conference on Robotics and Automation (ICRA) , 2019

    work page 2019

  44. [44]

    Son and B

    D. Son and B. Kim. Local object crop collision network for efficient simulation of non-convex objects in gpu-based simulators. arXiv preprint arXiv:2304.09439, 2023

    Pith/arXiv arXiv 2023

  45. [45]

    Featherstone

    R. Featherstone. Rigid Body Dynamics Algorithms . Springer-Verlag, 2007

    work page 2007

  46. [46]

    Karpathy

    A. Karpathy. nanoGPT, 2023. URL https://github.com/karpathy/nanoGPT

    work page 2023

  47. [47]

    Radford, J

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019

    work page 2019

  48. [48]

    Sutskever, O

    I. Sutskever, O. Vinyals, and Q. V . Le. Sequence to sequence learning with neural networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems - V olume 2, 2014

    work page 2014

  49. [49]

    Shinbrot, C

    T. Shinbrot, C. Grebogi, J. Wisdom, and J. A. Yorke. Chaos in a double pendulum. American Journal of Physics, 1992

    work page 1992

  50. [50]

    Hutter, C

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V . Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, and M. Hoepflinger. Any- mal - a highly mobile and dynamic quadrupedal robot. In 2016 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS) , 2016

    work page 2016

  51. [51]

    Schulman, F

    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

  52. [52]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State. Isaac gym: High performance gpu-based physics simula- tion for robot learning, 2021. 12 Neural Robot Dynamics: Appendix Appendix Contents A Additional NeRD Details 14 B Additional Training Details and Hyperparameters 15 C Addit...

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.