Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Sym2Real: Symbolic Dynamics with Residual Learning for Data-Efficient Adaptive Control

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Sym2Real claims that roughly ten real-world trajectories, built on a symbolic-regression base from low-fidelity simulation plus a small neural residual, are enough to control a quadrotor and a racecar across out-of-distribution shifts and s

desk verdict The sim-first symbolic-regression-plus-residual pipeline is a genuinely useful idea, but the paper's headline real-world claim is not yet supported by the evidence it reports. read the letter →

arxiv 2509.15412 v2 pith:QVFNWXSR submitted 2025-09-18 cs.RO cs.SYeess.SY

Easop Lee , Samuel A. Moore , Boyuan Chen This is my paper
classification cs.ROcs.SYeess.SY
keywords symbolicregressionresiduallearningmodelpredictivecontrolsim-to-realtransferadaptivequadrotorracecardata-efficient
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a robot's dynamics can be learned for real-world control with roughly ten trajectories by splitting the problem in two: symbolic regression extracts the core physics equations from a deliberately low-fidelity, noiseless simulator, and then a small neural-network residual, trained on just a handful of real trajectories, corrects the sim-to-real gap. The authors demonstrate the approach on a quadrotor and a racecar across six simulated out-of-distribution shifts and five real-world conditions, including added mass, center-of-mass offset, wind, and steering bias. Why the claim matters: standard model-learning pipelines need either large datasets or extensive simulation tuning, and Sym2Real argues neither is necessary if the base physics is shared across domains. A sympathetic reader would take away that decomposing a dynamics model into a symbolic core plus a learned residual is a viable route to few-shot adaptation for low-level control.

What carries the argument

The central object is the decomposed dynamics model f(s,a) = f_SR(s,a) + f_res(s,a), where f_SR is a symbolic equation fitted per state dimension on clean low-fidelity simulation data and f_res is a small multilayer perceptron trained to minimize the L2 prediction error on real transitions subject to a regularization penalty that keeps the residual small. The load-bearing mechanism is that shared physics survives in a low-fidelity simulator, so the symbolic base gives safe zero-shot performance and the residual only has to capture domain-specific discrepancies such as mass offsets, wind, and actuator latency. The controller is MPPI (Model Predictive Path Integral), which queries the model au

What would settle it

Collect a held-out set of real-world transitions after adaptation and compare the one-step prediction error of the residual-augmented model against the frozen symbolic base; if the residual does not reduce prediction error on this held-out set, the adaptation claim fails. Alternatively, run the pipeline on a platform where the default simulator file is deliberately corrupted (e.g., motor gains halved), so that zero-shot symbolic deployment cannot hold the quadrotor stable long enough to collect any real data—this would violate the stated minimum-performance assumption.

Watch

Extended reading notes

Core claim

Sym2Real's central discovery is that symbolic regression, normally too noise-sensitive for real-world data, becomes a practical base model when fitted on noise-free low-fidelity simulation, and that the remaining domain gap is small and structured enough to be absorbed by a lightweight neural residual trained on one to five real trajectories. The pipeline is: collect under twenty simulated trajectories in a default-parameter simulator, fit one symbolic equation per state dimension with symbolic regression, deploy this frozen symbolic model under sampling-based MPC, then collect one target-domain episode at a time and train an MLP residual with an L2 prediction loss plus a regularization term

Load-bearing premise

The default low-fidelity simulator, with no injected effects or parameter identification, is accurate enough that the symbolic base model can stabilize the real system long enough to collect the first target-domain trajectory.

Editorial extensions

If this is right

  • With fewer than ten real trajectories, the symbolic-base-plus-neural-residual combination reduces tracking error under mass, center-of-mass, wind, and steering-offset disturbances on both tested platforms.
  • Directly fine-tuning the symbolic model on real-world data removes key action dependencies (e.g., thrust) and can make control unsafe, whereas residual learning preserves the base structure.
  • Symbolic base models generalize zero-shot under state-action coverage shifts where neural base models fail—90% vs 20% success for the racecar and 100% vs 30% for the quadrotor in the paper's tests.
  • The symbolic model's roughly 4x faster inference relative to a neural ensemble allows more candidate rollouts per control cycle under the same real-time budget.
  • The method claims to require no domain randomization, no simulator parameter tuning, and no hand-derived equations, only a standard robot description file and a generic physics engine.
  • The residual training is regularized toward small outputs, which keeps the learned correction from overpowering the physics-based base model during adaptation.

Reading between the lines

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

  • A testable extension is to replace the symbolic base with any compact, interpretable regressor for systems whose physics are not easily expressible, leaving symbolic regression for platforms where it excels; the paper only compares symbolic versus neural bases.
  • The residual MLP likely absorbs actuator delay, communication latency, and thrust-conversion error implicitly; one could inject known latency into simulation and measure whether the residual magnitude shrinks, isolating what the residual encodes.
  • The paper's 'minimum performance assumption' implies a bootstrapping threshold: platforms where the default simulator is grossly wrong (e.g., an incorrect motor model) may fail before adaptation begins, suggesting an extension where the base model is first validated under a safe, conservative controller.
  • The real-world evidence covers only two specific platforms and five scenarios, and 'no expert knowledge' is better read as 'no simulator tuning'—task costs and control parameters are still hand-specified.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Sym2Real, a two-stage framework for data-efficient adaptive control. First, a symbolic-regression base model is fit to trajectories from a deliberately low-fidelity MuJoCo simulator (no simulator tuning, no domain randomization). Second, a lightweight neural-network residual is trained on a small number of real-world trajectories to compensate for the sim-to-real gap, and the combined model is deployed with an MPPI controller. The authors claim that using about 10 trajectories, the method achieves robust control of a Crazyflie quadrotor and a MuSHR racecar in the real world, without expert knowledge or simulation tuning, and that it adapts consistently across 6 out-of-distribution sim2sim scenarios and 5 real-world conditions. The manuscript reports several baselines corresponding to different base/residual model combinations, and includes both simulation and hardware experiments.

Significance. If the central claim holds, the contribution is significant: it combines symbolic regression with residual learning to make model-based MPC few-shot adaptable on real robots, and it does so using only a low-fidelity simulator as a black-box oracle. The idea that shared physics can be extracted from simplified simulation and then specialized with a small neural residual is well motivated, and the comparison of four base+residual combinations addresses an important design question. The paper also explicitly demonstrates that direct symbolic regression on noisy real-world data degrades, which is a useful negative result. However, the headline quantitative claim is not yet supported by the evidence presented: the real-world experiments are reported only qualitatively, and the paper does not provide code, raw data, or per-trial results. The strengths of the manuscript are the breadth of the experimental scenarios and the clear two-stage pipeline, but the current evaluation is not sufficient to establish the claimed robustness in a falsifiable way.

major comments (4)
  1. [Section VIII / V-C] The performance metric defined in Section V-C, average Euclidean position error, is never reported for any real-world scenario. Section VIII presents only qualitative trajectory plots (Figs. 7 and 8) and text statements such as 'residual learning consistently improved control performance.' Fig. 8 says results are averaged over three runs, but no error bars, per-trial values, or final position errors are given. Because the central claim is robust real-world control with about 10 trajectories and consistent improvement across five conditions, the evidence as presented is not falsifiable: the reader cannot compute the defined metric, cannot verify the improvement size, and cannot separate scenario-specific effects from general adaptation. Please add a quantitative table/report of the defined error for every real-world scenario and baseline, with sample counts, standard deviations, and prefe
  2. [Section IV-B] The 'minimum performance assumption' — that the base SR model is accurate enough to safely collect real-world data — is acknowledged but never quantified. Section VIII states that the SR model was 'accurate enough' for real-world deployment, but no threshold, metric, or measurement is provided. This is load-bearing because if the sim-to-real gap exceeds this unstated threshold, the pipeline cannot start, as the authors themselves note. Please define an operational criterion (e.g., bounded tracking error over a safe trajectory, or a safety-contract test) and report the measured zero-shot errors for each platform and scenario against that criterion. Without this, the safety premise underlying the entire adaptation process remains an unvalidated assumption.
  3. [Abstract / Section I / Section VIII] The advertised data budget is inconsistent and never precisely specified for real-world experiments. The abstract says 'about 10 trajectories,' Section I says 'we adapt in <5 trajectories across 5 scenarios,' and Section VII reports 'fewer than 10 additional trajectories' for sim2sim. Section VIII does not state the number of trajectories used per real-world condition, the criterion for stopping ('no meaningful improvement' in Section IV-B is not formalized), or whether the budget includes safety/reset episodes. The data-efficiency claim requires a precise per-scenario budget and an operational termination rule. Please report these for each of the five real-world scenarios.
  4. [Section VIII] The manuscript states that 'SR residual baselines exhibited unstable performance fluctuations during adaptation that resulted in hardware damages,' but does not specify which scenarios, how many trials were lost, which hardware was damaged, or how these events affected the averaged results. Since some reported results are averaged over three runs, a hardware-damage-related early termination could bias the comparison. Please provide complete trial accounting for all baselines and scenarios, including failed/damaged runs.
minor comments (6)
  1. [Section IV-A2 / Table I] The representative symbolic equations are shown with truncated terms. Full expressions, or a supplementary file, would be needed for reproducibility and for verifying the claim that action variables appear in the relevant state updates.
  2. [Section IV-B / Eq. (5)] The residual regularization weight lambda, the MPC cost weights, the MPPI sample count and horizon, the MLP architecture, and the PySR settings are not reported in a single table. This would help readers assess the 'no tuning' claim and reproduce the results.
  3. [Section V-B] The thrust conversion is an approximate linear mapping based on firmware values, and the paper acknowledges this as a source of error. Please quantify the uncertainty in this mapping or describe how the residual learning compensates for it, since it is one of the domain gaps being bridged.
  4. [Section I / Abstract] The trajectory-count wording should be made consistent ('about 10', '<5', 'fewer than 10'). This is not just a presentation issue because the readers and referees need to know the exact budget for each scenario.
  5. [Section IV-C] The comparison with model-free RL is described in one sentence ('crashed the drone'), without experimental detail or quantitative support. If this comparison is retained, please provide more details; otherwise it could be removed or softened.
  6. [General] The paper would benefit from releasing code and raw experimental data (trajectories, per-trial metrics) at the project website. This is especially important given the qualitative nature of the current real-world results.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central sim2real claim is evaluated on independent real hardware, and the residual-learning step is an explicitly data-driven adaptation rather than a disguised prediction.

full rationale

The derivation chain is self-contained in the relevant sense. Stage I fits SR equations to low-fidelity simulation trajectories (Eq. 3) and Stage II fits an NN residual to target-domain data (Eqs. 4-5); the evaluation in Sections VII-VIII measures closed-loop tracking on the target platform, which is independent of the fitting constants used in simulation. The sim2sim experiments compare SR+NN against baselines on OOD scenarios, and the sim2real experiments deploy the adapted model on physical Crazyflie and MuSHR platforms, so the central claim does not reduce to its inputs by construction. The few self-citations ([10], [22], [38]) appear in related-work and future-work contexts and are not load-bearing. Two non-circular concerns are noted for completeness: (i) the 'minimum performance assumption' (Section IV-B) is acknowledged but never quantitatively defined or validated, and (ii) the real-world results in Section VIII provide qualitative trajectory plots and averaged statements but do not report the paper's own position-error metric defined in Section V-C, making the headline ~10-trajectory robustness claim hard to falsify. These are reporting/verifiability issues, not circularity: the fitted residual is used to adapt, not to 'predict' the evaluation outcome. Hence score 1.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central method adds no new physical entity; it relies on a composite base+residual model. The free parameters are largely unreported controller and training knobs, which raises reproducibility burden but not circularity.

free parameters (6)
  • lambda (residual regularization weight) = not reported
    Eq. 5: lambda controls the trade-off between prediction error and residual magnitude; hand-set, no sensitivity study.
  • MPC tracking cost weights alpha_p, alpha_r, alpha_v, alpha_omega = not reported
    Eq. 6: these weights define the task objective for MPPI and are selected per task; no values or tuning procedure given.
  • MPPI sample count N and horizon H = N=1024, H=0.8s reported only for the speed comparison
    Controller hyperparameters affect stability and performance; not specified for the adaptation experiments.
  • PySR expression length cap and operator set = length cap 85; operators {+,-,*,cos,sin}; 5 search iterations
    Section IV-A2: these settings constrain the model class; authors say no extensive search but the choices are manual.
  • Residual MLP architecture and training hyperparameters = not reported
    Layer sizes, activation, learning rate, optimizer, and number of gradient steps are absent, so the residual learner cannot be reproduced.
  • Thrust conversion linear mapping = approximate linear mapping from firmware values
    Section V-B: a manual unit conversion from Newtons to 16-bit thrust; its error is a domain gap the residual is expected to absorb.
assumptions (6)
  • domain assumption Shared core physics: low-fidelity simulation preserves the essential dynamics of the real platform.
    Section IV-A1 and IV-B: without this, the SR base model would be unsafe for real-world data collection and residual adaptation.
  • domain assumption MuJoCo with default MJCF files and no injected noise is an adequate oracle for symbolic regression.
    Section IV-A: low-fidelity means default robot description files, noiseless observations, no latency; the base equations are assumed to capture true action-to-state dependencies.
  • standard math The state is Markovian: next state depends only on current state and action.
    Section III Eq. 1: the learned model f: S x A -> S assumes one-step predictive sufficiency; delays and histories are not modeled.
  • domain assumption A known task cost function c(s,a) is available.
    Section III Eq. 2 and Section V-C Eq. 6: MPPI requires a hand-specified objective; the cost weights are treated as given rather than learned.
  • domain assumption Sampling-based MPPI with the learned model is sufficient to stabilize the platform.
    Section III and IV-C: controller choice is justified by a single comparative experiment (RL crashed, MPC hovered), not by an analysis.
  • domain assumption The PySR operator set and expression-length cap are expressive enough to represent the core dynamics.
    Section IV-A2: if the true dynamics require operators outside {+,-,*,cos,sin} or longer expressions, the base model will miss structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sym2Real: Symbolic Dynamics with Residual Learning for Data-Efficient Adaptive Control." pith.science (2026). https://pith.science/paper/QVFNWXSR

@misc{pith2026250915412,
  author       = {Pith},
  title        = {Pith review of: Sym2Real: Symbolic Dynamics with Residual Learning for Data-Efficient Adaptive Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QVFNWXSR}},
  note         = {Machine review of arXiv:2509.15412}
}
read the original abstract

We present Sym2Real, a fully data-driven framework for highly data-efficient adaptation of low-level controllers. Although symbolic regression is data-efficient, its role in real-world control has been limited due to its sensitivity to measurement noise, which corrupts the equations and leads to model degradation when fitted directly on real-world data. Sym2Real addresses this limitation by 1) learning first from low-fidelity simulation, where noise-free trajectories allow symbolic regression to identify the underlying dynamics, and 2) using a small amount of real-world data for targeted residual adaptation to bridge the sim-to-real gap. Using only about 10 trajectories, we achieve robust control of both a quadrotor and a racecar in the real world, without expert knowledge or simulation tuning. Through experimental validation on both platforms, we demonstrate consistent data-efficient adaptation across 6 out-of-distribution sim2sim scenarios and successful sim2real transfer across 5 real-world conditions. More information can be found at http://generalroboticslab.com/Sym2Real

Figures

Figures reproduced from arXiv: 2509.15412 by the authors.

Figure 1
Figure 1. Sym2Real enables robust and adaptive real-world control using a small amount of simulation data (light blue background) and real-world data (light orange background). trial-and-error process that presupposes dynamics knowledge in the target domain, which offers limited reusability across platforms or tasks [6], [7]. Together, these challenges collec￾tively impede the development of robust and adaptive low￾level data… view at source ↗
Figure 2
Figure 2. Sym2Real achieves data-efficient adaptive control through a two-stage process. (a) Simplified Environment: In Stage I, the SR dynamics model extracts core equations from a simplified low-fidelity environment, where trajectories, including high-speed drifts or flips that would trigger safety shutdowns in reality, can be safely simulated to provide rich state–action coverage from only a few runs. (b) Target Environmen… view at source ↗
Figure 3
Figure 3. SR vs. NN base model efficiency. (a) A single step model inference speed of the base models. (b) The data efficiency of the SR model in the simplified simulation environment. B. Real-World Hardware Configuration External motion capture (MoCap) data was streamed at 100 Hz to provide pose measurements. Crazyflie used only position from MoCap, while orientation and angular rate were from the onboard IMU. MuSHR racecar … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Residual adaptation performance. Episode cost during online adaptation for (a-f) the Crazyflie quadrotor under various disturbances and (g-i) the MuSHR racecar across scenarios. Results are averaged over three seeds, with shaded regions indicating standard deviation. T…
Figure 6
Figure 6. Figure 6: Experiment setup for zero-shot generalization under OOD state–action coverage. (a) The MuSHR racecar was trained with initial positions restricted to a narrow range (purple), and tested from disjoint start states (yellow). (b) The quadrotor was trained to remain hoveri…
Figure 5
Figure 5. Figure 5: Sim2sim adaptation performance of SR*+NN. Few-shot residual adaptation improves both model prediction and control compared to direct SR transfer: (a-c) shows Crazyflie quadrotor adaptation and (d-f) shows MuSHR racecar adaptation. Dotted black lines represent target re…
Figure 7
Figure 7. Figure 7: Sim2real Crazyflie 2.1 quadrotor adaptation. (a) Plain sim2real transfer, (b) added mass (one penny), (c) added mass (two pennies), and (d) center-of-mass offset. In (a) and (b), the quadrotor was tasked with hovering, while in (d) it was tasked to follow a circular tr…
Figure 8
Figure 8. Figure 8: Sim2real MuSHR racecar adaptation. (a) Plain sim2real transfer and (b) steering-offset setting. Performance of all baselines and Sym2Real (i.e., SR* + NN) across trials of few-shot adaptation, averaged over three runs is reported. SR* + NN residual adaptation improved …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 6 linked inside Pith

  1. [1]

    Champion-level drone racing using deep reinforce- ment learning,

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Champion-level drone racing using deep reinforce- ment learning,”Nature, vol. 620, no. 7976, pp. 982–987, 2023

    2023

  2. [2]

    Learning robust control policies for end-to- end autonomous driving from data-driven simulation,

    A. Amini, I. Gilitschenski, J. Phillips, J. Moseyko, R. Banerjee, S. Karaman, and D. Rus, “Learning robust control policies for end-to- end autonomous driving from data-driven simulation,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 1143–1150, 2020

    2020

  3. [3]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,”Science robotics, vol. 5, no. 47, p. eabc5986, 2020

    2020

  4. [4]

    Adapt- ing rapid motor adaptation for bipedal robots,

    A. Kumar, Z. Li, J. Zeng, D. Pathak, K. Sreenath, and J. Malik, “Adapt- ing rapid motor adaptation for bipedal robots,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 1161–1168

    2022

  5. [5]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in2017 IEEE/RSJ international con- ference on intelligent robots and systems (IROS). IEEE, 2017, pp. 23–30

    2017

  6. [6]

    What matters in learning a zero-shot sim-to-real rl policy for quadrotor control? a comprehensive study,

    J. Chen, C. Yu, Y . Xie, F. Gao, Y . Chen, S. Yu, W. Tang, S. Ji, M. Mu, Y . Wu, H. Yang, and Y . Wang, “What matters in learning a zero-shot sim-to-real rl policy for quadrotor control? a comprehensive study,” IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 7134–7141, 2025

    2025

  7. [7]

    Using simulation optimization to improve zero-shot policy transfer of quadrotors,

    S. Gronauer, M. Kissel, L. Sacchetto, M. Korte, and K. Diepold, “Using simulation optimization to improve zero-shot policy transfer of quadrotors,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 10 170–10 176

    2022

  8. [8]

    Symbolic-model-based reinforcement learning,

    P.-A. Kamienny and sylvain lamprier, “Symbolic-model-based reinforcement learning,” inNeurIPS 2022 AI for Science: Progress and Promises, 2022. [Online]. Available: https://openreview.net/ forum?id=yeF6cyYU7W

    2022

Show all 38 references
  1. [9]

    Model-based policy optimization using symbolic world model,

    A. Gorodetskiy, K. Mironov, and A. Panov, “Model-based policy optimization using symbolic world model,” in2024 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS), 2024, pp. 664–669

    2024

  2. [10]

    Automated global analysis of experimental dynamics through low-dimensional linear embeddings,

    S. A. Moore, B. P. Mann, and B. Chen, “Automated global analysis of experimental dynamics through low-dimensional linear embeddings,” arXiv preprint arXiv:2411.00989, 2024

    2024 arXiv

  3. [11]

    Deep rein- forcement learning in a handful of trials using probabilistic dynamics models,

    K. Chua, R. Calandra, R. McAllister, and S. Levine, “Deep rein- forcement learning in a handful of trials using probabilistic dynamics models,”Advances in neural information processing systems, vol. 31, 2018

    2018

  4. [12]

    Low-level control of a quadrotor with deep model- based reinforcement learning,

    N. O. Lambert, D. S. Drew, J. Yaconelli, S. Levine, R. Calandra, and K. S. Pister, “Low-level control of a quadrotor with deep model- based reinforcement learning,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 4224–4230, 2019

    2019

  5. [13]

    Pilco: a model-based and data-efficient approach to policy search,

    M. P. Deisenroth and C. E. Rasmussen, “Pilco: a model-based and data-efficient approach to policy search,” inProceedings of the 28th International Conference on International Conference on Machine Learning, ser. ICML’11. Madison, WI, USA: Omnipress, 2011, p. 465–472

    2011

  6. [14]

    Neural network dynamics for model-based deep reinforcement learning with model- free fine-tuning,

    A. Nagabandi, G. Kahn, R. S. Fearing, and S. Levine, “Neural network dynamics for model-based deep reinforcement learning with model- free fine-tuning,” in2018 IEEE international conference on robotics and automation (ICRA). IEEE, 2018, pp. 7559–7566

    2018

  7. [15]

    Sindy-rl: interpretable and efficient model-based reinforcement learning,

    N. Zolman, U. Fasel, J. N. Kutz, and S. L. Brunton, “Sindy-rl: interpretable and efficient model-based reinforcement learning,”arXiv preprint arXiv:2403.09110, 2024

    2024 arXiv

  8. [16]

    Physics-informed neural net- work for quadrotor dynamical modeling,

    W. Gu, S. Primatesta, and A. Rizzo, “Physics-informed neural net- work for quadrotor dynamical modeling,”Robotics and Autonomous Systems, vol. 171, p. 104569, 2024

    2024

  9. [17]

    Physics-informed dyna-style model-based deep reinforcement learning for dynamic control,

    X.-Y . Liu and J.-X. Wang, “Physics-informed dyna-style model-based deep reinforcement learning for dynamic control,”Proceedings of the Royal Society A, vol. 477, no. 2255, p. 20210618, 2021

    2021

  10. [18]

    Symbolic regression via genetic pro- gramming,

    D. Augusto and H. Barbosa, “Symbolic regression via genetic pro- gramming,” inProceedings. Vol.1. Sixth Brazilian Symposium on Neural Networks, 2000, pp. 173–178

    2000

  11. [19]

    Distilling free-form natural laws from experimental data,

    M. Schmidt and H. Lipson, “Distilling free-form natural laws from experimental data,”science, vol. 324, no. 5923, pp. 81–85, 2009

    2009

  12. [20]

    Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients,

    B. K. Petersen, M. Landajuela, T. N. Mundhenk, C. P. Santiago, S. K. Kim, and J. T. Kim, “Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients,” inProc. of the International Conference on Learning Representations, 2021

    2021

  13. [21]

    Interpretable machine learning for science with pysr and symbolicregression. jl,

    M. Cranmer, “Interpretable machine learning for science with pysr and symbolicregression. jl,”arXiv preprint arXiv:2305.01582, 2023

    2023 arXiv

  14. [22]

    Symmatika: Structure-aware symbolic discovery,

    M. Scherk and B. Chen, “Symmatika: Structure-aware symbolic discovery,” 2025. [Online]. Available: https://arxiv.org/abs/2507.03110

    2025 arXiv

  15. [23]

    Distilling reinforcement learning policies for interpretable robot locomotion: Gradient boosting machines and sym- bolic regression,

    F. Acero and Z. Li, “Distilling reinforcement learning policies for interpretable robot locomotion: Gradient boosting machines and sym- bolic regression,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 6840–6847

    2024

  16. [24]

    Gaussian processes and reinforcement learning for identification and control of an au- tonomous blimp,

    J. Ko, D. J. Klein, D. Fox, and D. Haehnel, “Gaussian processes and reinforcement learning for identification and control of an au- tonomous blimp,” inProceedings 2007 IEEE International Conference on Robotics and Automation, 2007, pp. 742–747

    2007

  17. [25]

    Data-driven mpc for quadrotors,

    G. Torrente, E. Kaufmann, P. F ¨ohn, and D. Scaramuzza, “Data-driven mpc for quadrotors,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3769–3776, 2021

    2021

  18. [26]

    Scalable deep kernel gaussian process for vehicle dynamics in autonomous racing,

    J. Ning and M. Behl, “Scalable deep kernel gaussian process for vehicle dynamics in autonomous racing,” inConference on Robot Learning. PMLR, 2023, pp. 761–773

    2023

  19. [27]

    Knode-mpc: A knowledge- based data-driven predictive control framework for aerial robots,

    K. Y . Chee, T. Z. Jiahao, and M. A. Hsieh, “Knode-mpc: A knowledge- based data-driven predictive control framework for aerial robots,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2819–2826, 2022

    2022

  20. [28]

    How to learn and generalize from three minutes of data: Physics-constrained and uncertainty-aware neural stochastic differential equations,

    F. Djeumou, C. Neary, and ufuk topcu, “How to learn and generalize from three minutes of data: Physics-constrained and uncertainty-aware neural stochastic differential equations,” in7th Annual Conference on Robot Learning, 2023

    2023

  21. [29]

    Sim- to-real of soft robots with learned residual physics,

    J. Gao, M. Y . Michelis, A. Spielberg, and R. K. Katzschmann, “Sim- to-real of soft robots with learned residual physics,”IEEE Robotics and Automation Letters, 2024

    2024

  22. [30]

    Learning to walk from three minutes of real-world data with semi-structured dynamics models,

    J. Levy, T. Westenbroek, and D. Fridovich-Keil, “Learning to walk from three minutes of real-world data with semi-structured dynamics models,” in8th Annual Conference on Robot Learning, 2024

    2024

  23. [31]

    DATT: Deep adaptive trajectory tracking for quadrotor control,

    K. Huang, R. Rana, A. Spitzer, G. Shi, and B. Boots, “DATT: Deep adaptive trajectory tracking for quadrotor control,” in7th Annual Conference on Robot Learning, 2023

    2023

  24. [32]

    Neural-fly enables rapid learning for agile flight in strong winds,

    M. O’Connell, G. Shi, X. Shi, K. Azizzadenesheli, A. Anandkumar, Y . Yue, and S.-J. Chung, “Neural-fly enables rapid learning for agile flight in strong winds,”Science Robotics, vol. 7, no. 66, p. eabm6597, 2022

    2022

  25. [33]

    ASID: Active exploration for system identification in robotic manipulation,

    M. Memmel, A. Wagenmaker, C. Zhu, D. Fox, and A. Gupta, “ASID: Active exploration for system identification in robotic manipulation,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  26. [34]

    Finetuning offline world models in the real world,

    Y . Feng, N. Hansen, Z. Xiong, C. Rajagopalan, and X. Wang, “Finetuning offline world models in the real world,” inProceedings of the 7th Conference on Robot Learning (CoRL), 2023

    2023

  27. [35]

    Aggressive driving with model predictive path integral control,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Aggressive driving with model predictive path integral control,” in 2016 IEEE International Conference on Robotics and Automation (ICRA), 2016, pp. 1433–1440

    2016

  28. [36]

    On the benefits of gpu sample-based stochastic predictive controllers for legged locomotion,

    G. Turrisi, V . Modugno, L. Amatucci, D. Kanoulas, and C. Semini, “On the benefits of gpu sample-based stochastic predictive controllers for legged locomotion,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, Oct. 2024, p. 13757–13764

    2024

  29. [37]

    Mushr: A low-cost, open-source robotic racecar for education and research,

    S. S. Srinivasa, P. Lancaster, J. Michalove, M. Schmittle, C. Summers, M. Rockett, R. Scalise, J. R. Smith, S. Choudhury, C. Mavrogiannis, et al., “Mushr: A low-cost, open-source robotic racecar for education and research,”arXiv preprint arXiv:1908.08031, 2019

    1908 arXiv

  30. [38]

    Automated discovery of continuous dynamics from videos,

    K. Huang, D. H. Cho, and B. Chen, “Automated discovery of continuous dynamics from videos,” 2024. [Online]. Available: https://arxiv.org/abs/2410.11894

    2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.