Pith. sign in

REVIEW 4 major objections 5 minor 52 references

DiffCoTune: Differentiable Co-Tuning for Cross-domain Robot Control

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Co-tuning the simulator's physics parameters and the controller's weights, using one real-world rollout per iteration, transfers robot controllers with at most five trials.

desk verdict A useful, honest extension of DiffTune with a real identifiability gap in the beta-update and thinner empirical reporting than the few-trial claim requires. read the letter →

arxiv 2505.24068 v1 pith:ALSABK3F submitted 2025-05-29 cs.RO

Lokesh Krishna , Sheng Cheng , Junheng Li , Naira Hovakimyan , Quan Nguyen This is my paper
classification cs.RO
keywords differentiablesimulationsim-to-realtransfercontrollertuningsystemidentificationalternatingoptimizationleggedlocomotiongradient-basedmodel-basedcontrol
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

The paper proposes DiffCoTune, a gradient-based procedure that tunes a robot controller's parameters together with the parameters of its differentiable simulator so that performance in the deployment domain improves within a few real-world rollouts. The central claim is that coupling system identification and controller adaptation in an alternating optimization—rather than first identifying the model and then tuning the controller—is what makes the transfer work under model mismatch. The authors test this on cart-pole stabilization, quadruped velocity tracking, and bipedal jumping, with controllers ranging from LQR to MPC to deep neural network policies, including two hardware experiments. If the claim holds, it provides a generic alternative to ad-hoc manual tuning for sim-to-real transfer.

What carries the argument

The load-bearing mechanism is the alternating co-tuning update. In each outer iteration, one rollout of the current controller on the target system is collected; a system-identification step then adjusts $\beta$ with $\theta$ held fixed by minimizing $J_{\mathrm{sysId}}(\theta_l,\beta,\{x^{\mathrm{sys}}_t\}_l)$, and a task step adjusts $\theta$ with the new $\beta$ held fixed by minimizing $J_{\mathrm{task}}(\theta,\beta_{l+1})$. Because both steps differentiate through the simulator rollout, a handful of deployment rollouts steers both the model and the controller. The paper finds that this split-alternate scheme is robust across systems and controllers, whereas the combined weighted objective $w_{\mathrm{task}}J_{\mathrm{task}}+w_{\mathrm{sysId}}J_{\mathrm{sysId}}$ diverges under model mismatch.

What would settle it

Take a deployment platform whose dominant mismatch is deliberately outside the tunable parameter set—for example, actuator delay or link flexibility while $\beta$ contains only link masses—and run the method with held-out deployment rollouts. If the held-out closed-loop prediction error does not decrease while task loss improves, or if the advantage over controller-only tuning disappears, the central claim that co-tuning of $\beta$ drives the transfer would be falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that sim-to-real transfer can be solved as a co-tuning problem. Given a nominal controller $\pi(\cdot;\theta)$ and a differentiable simulator $x_{t+1}=f(x_t,u_t;\beta)$, each iteration collects one rollout of the current controller in the target system, updates the simulator parameters $\beta$ by minimizing a closed-loop system-identification loss $J_{\mathrm{sysId}}$ that compares model and system trajectories under the same controller, and then updates the controller parameters $\theta$ by minimizing the simulated task loss $J_{\mathrm{task}}$ in the updated simulator. The paper argues that the resulting Split-Alternate update is stable across systems and controllers without task-specific hyperparameter tuning, and that co-tuning outperforms tuning the controller alone (DiffTune), a one-shot sysId-then-tune pipeline, and Bayesian optimization. Reported results include improved tracking with at most five target rollouts: a quadruped that falls without tuning completes forward-and-yaw tracking after tuning, and bipedal jump length improves from 20 cm for the untuned controller to 34 cm with co-tuning on hardware, against a synthesized 40 cm reference.

Load-bearing premise

The load-bearing premise is that the simulator's tunable parameters are rich enough to match the real system's dynamics: if the true mismatch lies outside the parameterized model class, the $\beta$-update cannot close the sim-to-real gap and the controller tunes against the wrong model.

Editorial extensions

If this is right

  • With at most five deployment rollouts, co-tuning improves performance for LQR, PD, MPC, and DNN policies on cart-pole, quadruped, and biped systems, under sim-to-sim and sim-to-real transfer.
  • Under increasing mass mismatch (15%, 30%, 45%), co-tuning matches or outperforms controller-only gradient tuning, with the gap widening at larger mismatch; at 60% mismatch both methods degrade.
  • Iteratively collecting data with the tuned controller and co-tuning on each rollout beats collecting all rollouts upfront and then tuning (sysId-then-tune).
  • Gradient-based co-tuning reduces task loss within about 10 target trials, much faster than Bayesian optimization, which needed about 90 trials on the cart-pole LQR task.
  • A single algorithmic choice, the Split-Alternate update with fixed learning rate and epochs, transfers across all tested systems without per-task hyperparameter adjustment, while the Combined update is fragile under mismatch.

Reading between the lines

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

  • (Editorial extension) The assumption that $\beta$ can represent the real system suggests that practical deployments should include residual dynamics terms in $\beta$; the paper lists DNN residual co-tuning as future work, so the method's universality across unmodeled effects remains untested.
  • (Editorial extension) Because both methods degrade at extreme 60% mismatch, the method is best understood as local adaptation around a nominal controller that already stabilizes the system, not as a replacement for control synthesis.
  • (Editorial extension) A direct diagnostic for practitioners would be to track held-out closed-loop prediction error: if task loss improves while $J_{\mathrm{sysId}}$ stays flat on fresh rollouts, the gains may come from exploiting simulator artifacts rather than from closing the sim-to-real gap.
  • (Editorial extension) The same alternating gradient loop could in principle tune non-differentiable controllers through implicit differentiation or zero-order gradient estimates, but the paper does not demonstrate this.
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 / 5 minor

Summary. The paper proposes DiffCoTune, a framework for co-tuning the parameters of a differentiable simulator and a differentiable controller so that a nominal controller transfers better to a deployment domain. The authors formulate two update strategies: Combined, which minimizes a weighted sum of a task objective and a closed-loop system-identification objective, and Split-Alternate, which alternates between fitting the simulator to a closed-loop system rollout and tuning the controller in the updated simulator. Experiments cover cart-pole stabilization, quadruped and biped locomotion, and sim-to-sim and sim-to-real transfer, comparing against DiffTune, Bayesian optimization, and a sysId-then-tune baseline. The main claims are that the approach improves deployment performance within a few rollouts and scales to high-dimensional controllers with arbitrary complexity.

Significance. If the empirical claims hold, DiffCoTune is a useful addition to the controller auto-tuning literature, combining gradient-based system identification and controller adaptation without requiring the model to be perfectly identified. A notable strength is the breadth of the evaluation: the method is applied to model-based controllers (LQR, MPC, trajectory optimization) and to large DNN policies, and it is demonstrated on real quadruped and biped hardware. The paper also directly compares against relevant gradient-based and gradient-free baselines. However, the contribution is primarily empirical, and the current evidence has important gaps: no error bars or repeated trials, an inconsistency in the rollout budget, and an open question about whether the identified simulator parameters generalize after the controller is updated.

major comments (4)
  1. [§III, §IV.B, Table I] The beta-update in Eq. (8) minimizes the closed-loop tracking loss J_sysId under the current controller theta_l. The resulting beta_{l+1} can fit the closed-loop map of theta_l rather than the true system dynamics, and nothing in the paper bounds the error of this estimate after the controller is updated to theta_{l+1}. The sufficient-parameterization assumption stated in Section III is not satisfied in the experiments: Table I restricts beta to link masses in the quadruped and biped cases, leaving contact, friction, and actuation effects outside the model class. This gap is load-bearing for the claim of systematic transfer within a few trials. Please provide an analysis or an explicit experiment that validates beta generalization, for example by comparing model rollouts with the updated controller against system rollouts, or by including residual dynamics in the tunable parameter set.
  2. [Algorithm 1, §V] The paper states that the method uses a single rollout in the target system per iteration and claims improvement within fewer than five trials, but Algorithm 1 requires two system rollouts per iteration: one at line 3 to compute the sysId update and one at line 5 to evaluate J_task(theta_{l+1}, sys) for best-iterate selection. With the reported L=5, the procedure may use up to ten system rollouts, contradicting the stated rollout budget. Please clarify whether line 5 uses a new rollout or a surrogate, and if a new rollout is needed, restate the actual trial count and the claim of fewer than five trials.
  3. [§V, Figs. 7–8] The quantitative results are reported without error bars, multiple seeds, or repeated hardware trials; the quadruped and biped hardware results shown in Figs. 4 and 5 appear to be single runs. The claimed improvements (60% for the quadruped, 29.29% for the biped MPC case, 34 cm for the biped jump) are therefore not statistically supported. Please add repeated simulator runs with variance, repeated hardware trials if feasible, or clearly label single-trial results and temper the corresponding claims. In addition, the Adam learning rate and other optimizer settings for Alg. 2 and Alg. 3 are not reported, which makes the experiments difficult to reproduce.
  4. [§V.B.5, Fig. 7] The paper reports that the Combined variant consistently diverges under model uncertainty and attributes this to sensitivity of the objective landscape with the chosen weights, but no analysis is provided. This divergence is evidence that the beta-theta coupling is delicate, and the claim that Split-Alternate is immune to this issue is not established by the present ablations, since no weight-sensitivity or objective-landscape analysis is given for Split-Alternate either. Please either provide a formal or empirical explanation of why alternating optimization avoids the difficulty, or soften the claim that Split-Alternate is robust across systems and controllers.
minor comments (5)
  1. [Throughout] There are several typos, including 'paramterized' in Eq. (1), 'differntiable' in Section III, and 'intialize' in Algorithms 2 and 3; these should be corrected.
  2. [Fig. 1] The supplementary video link is shown as a placeholder 'link'; please replace it with the actual URL.
  3. [Eq. (7), §V] The weights w_task and w_sysId are introduced in Eq. (7) and then fixed to 1 in the experiments; please clarify whether these are intended as user-provided hyperparameters in general, and discuss the sensitivity of the Combined variant to them.
  4. [Table I] The row for the biped TO/PD case is difficult to parse: 'PD(g(x);theta) TO w/ varying models' does not clearly describe the controller structure or what theta and beta represent. Please expand the notation.
  5. [Section V.B.1] The sentence beginning 'A Auto-Differentiation' contains an article error; it should read 'An Auto-Differentiation'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the co-tuning updates and the empirical evaluation are independent, and self-citations are only used as baselines.

full rationale

The paper's derivation chain is self-contained: the simulator update (Eq. 8) minimizes the closed-loop system-identification loss J_sysId defined in Eq. (3), and the controller update (Eq. 9) minimizes the simulated task loss J_task defined in Eq. (5), with the actual deployment objective J_task(theta, sys) from Eq. (4) used only for evaluation in Algorithm 1. The beta-update is a fit to observed closed-loop rollouts, and the theta-update subsequently optimizes the controller in the fitted model; this is a standard alternating-optimization tuning procedure, not a prediction that is equal by construction to its inputs. The assumption in Section III that Eq. (1) is 'a sufficiently parameterized model to fit the real-world system to arbitrary accuracy' is a stated modeling condition, and any violation would weaken the empirical results rather than make them circular. The paper's self-citations, chiefly to DiffTune [27], are used for baseline comparisons and for pointing to prior quantitative studies, not as load-bearing evidence for the central co-tuning claim. Performance improvements are reported on the same deployment tasks used for tuning, which is ordinary controller tuning evaluation rather than a hidden reduction of the result to the input. No fitted parameter is renamed as a prediction, and no load-bearing premise is imported solely from the authors' prior work. Overall, the paper exhibits no identifiable circularity.

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

The central claim rests on the model class being expressive enough to capture the target system, differentiability of both the simulator and controller, and the local validity of gradient updates from a nominal controller. No new physical entities are introduced. Several hand-chosen hyperparameters (learning rate, max epochs, termination tolerance) are not fully disclosed.

free parameters (4)
  • Adam learning rate = not reported
    Hand-chosen, not specified in the paper; it significantly affects convergence and the reported performance improvements.
  • Maximum inner epochs K = 100
    Set to 100 for all experiments; the number of gradient updates in each tuning iteration is a design choice that impacts the results.
  • Termination tolerance = 1e-3
    Termination condition C stops when the objective change is below 1e-3 or when the objective increases by more than 1e-3, acting as a safeguard that selects the best iterate.
  • Combined weights w_task and w_sysId = 1 and 1
    Used only in the Combined variant; the paper notes this 'naive choice' exposes sensitivity to objective curvature and the variant diverges under model uncertainty.
assumptions (4)
  • domain assumption The discrete-time dynamics f(x,u;beta) is sufficiently parameterized to fit the real-world system to arbitrary accuracy.
    Section III states this assumption directly; if the model class cannot represent the true system, beta-updates cannot close the sim-to-real gap and theta-updates may optimize the wrong objective.
  • domain assumption Both the simulator and controller are differentiable, providing gradients with respect to states, inputs, and parameters.
    Section III defines differentiable simulator and controller; the entire gradient-based tuning framework depends on these gradients being available.
  • domain assumption A nominal controller synthesized for the nominal model is a sufficiently good starting point that local gradient updates improve the true task objective.
    Problem 1 starts from (beta_tilde, theta_tilde) and the method only searches locally; if the nominal controller is far from deployable, tuning is replaced by synthesis.
  • domain assumption The simulated task objective J_task(theta,beta) is a valid surrogate for the deployment task J_task(theta,sys) after beta-updates.
    The paper uses J_task on model rollouts to update theta, implicitly assuming that minimizing the model-based loss transfers to the real system once beta is updated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffCoTune: Differentiable Co-Tuning for Cross-domain Robot Control." pith.science (2026). https://pith.science/paper/ALSABK3F

@misc{pith2026250524068,
  author       = {Pith},
  title        = {Pith review of: DiffCoTune: Differentiable Co-Tuning for Cross-domain Robot Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALSABK3F}},
  note         = {Machine review of arXiv:2505.24068}
}
read the original abstract

The deployment of robot controllers is hindered by modeling discrepancies due to necessary simplifications for computational tractability or inaccuracies in data-generating simulators. Such discrepancies typically require ad-hoc tuning to meet the desired performance, thereby ensuring successful transfer to a target domain. We propose a framework for automated, gradient-based tuning to enhance performance in the deployment domain by leveraging differentiable simulators. Our method collects rollouts in an iterative manner to co-tune the simulator and controller parameters, enabling systematic transfer within a few trials in the deployment domain. Specifically, we formulate multi-step objectives for tuning and employ alternating optimization to effectively adapt the controller to the deployment domain. The scalability of our framework is demonstrated by co-tuning model-based and learning-based controllers of arbitrary complexity for tasks ranging from low-dimensional cart-pole stabilization to high-dimensional quadruped and biped tracking, showing performance improvements across different deployment domains.

Figures

Figures reproduced from arXiv: 2505.24068 by the authors.

Figure 1
Figure 1. Overview of the proposed automated co-tuning approach for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Phase portraits of progressive tuning iterates of DiffCoTune [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Sim-to-sim transfer of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Sim-to-real transfer of high-speed forward motion of {quadruped, DNN} on a Unitree Go1 quadruped. tracking task with desired base position commands in the world frame: p z t = 0.55 + 0.05 sin( 2πt T ), p x t = 0, and p y t = 0. Taking θ to be corresponding terms in dia…
Figure 6
Figure 6. Figure 6: Performance loss comparison of the proposed DiffCoTune [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 5
Figure 5. Figure 5: Comparative experimental snapshots of bipedal jumping [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: The performance loss along tuning iterations for tuning approaches over different [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Effect of varying magnitude (left) and nature (right) of [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 39 canonical work pages

  1. [17]

    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

  2. [1]

    Stochastic and robust mpc for bipedal locomotion: A comparative study on robustness and performance,

    A. Gazar, M. Khadiv, A. D. Prete, and L. Righetti, “Stochastic and robust mpc for bipedal locomotion: A comparative study on robustness and performance,” in2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids), pp. 61–68, 2021

    work page 2021

  3. [2]

    Adaptive-force-based control of dynamic legged locomotion over uneven terrain,

    M. Sombolestan and Q. Nguyen, “Adaptive-force-based control of dynamic legged locomotion over uneven terrain,”IEEE Transactions on Robotics, vol. 40, pp. 2462–2477, 2024

    work page 2024

  4. [3]

    Estimation of inertial parameters of manipulator loads and links,

    C. G. Atkeson, C. H. An, and J. M. Hollerbach, “Estimation of inertial parameters of manipulator loads and links,”The International Journal of Robotics Research, vol. 5, no. 3, pp. 101–119, 1986

    1986

  5. [4]

    Sim-to- real transfer of robotic control with dynamics randomization,

    X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel, “Sim-to- real transfer of robotic control with dynamics randomization,” in2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 3803–3810, 2018

    work page 2018

  6. [5]

    Learning Agile Robotic Locomotion Skills by Imitating Animals,

    X. B. Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine, “Learning Agile Robotic Locomotion Skills by Imitating Animals,” inProceedings of Robotics: Science and Systems, (Corvalis, Oregon, USA), July 2020

    work page 2020

  7. [6]

    Learning agile and dynamic motor skills for legged robots,

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter, “Learning agile and dynamic motor skills for legged robots,”Science Robotics, vol. 4, no. 26, p. eaau5872, 2019

    2019

  8. [7]

    Sim-to-real: Learning agile locomotion for quadruped robots,

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bo- hez, and V . Vanhoucke, “Sim-to-real: Learning agile locomotion for quadruped robots,” inProceedings of Robotics: Science and Systems, (Pittsburgh, Pennsylvania), June 2018

    work page 2018

Show all 52 references
  1. [8]

    Domain randomization via entropy maximization,

    G. Tiboni, P. Klink, J. Peters, T. Tommasi, C. D’Eramo, and G. Chal- vatzaki, “Domain randomization via entropy maximization,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  2. [9]

    Learning fast adapta- tion with meta strategy optimization,

    W. Yu, J. Tan, Y . Bai, E. Coumans, and S. Ha, “Learning fast adapta- tion with meta strategy optimization,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 2950–2957, 2020

    2020

  3. [10]

    P. A. Ioannou and J. Sun,Robust adaptive control. USA: Prentice-Hall, Inc., 1995

    1995

  4. [11]

    Hovakimyan and C

    N. Hovakimyan and C. Cao,L 1 Adaptive Control Theory. Society for Industrial and Applied Mathematics, 2010

    2010

  5. [12]

    Rma: Rapid motor adaptation for legged robots,

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

    2021

  6. [13]

    Continuous dynamic bipedal jumping via real-time variable-model optimization,

    J. Li, O. Kolt, and Q. Nguyen, “Continuous dynamic bipedal jumping via real-time variable-model optimization,” 2024

    2024

  7. [14]

    Ogmp: Oracle guided multi-mode policies for agile and versatile robot control,

    L. Krishna, N. Sobanbabu, and Q. Nguyen, “Ogmp: Oracle guided multi-mode policies for agile and versatile robot control,” 2024

    2024

  8. [15]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Pan, Z. Yi, G. Qu, K. Kitani, J. Hodgins, L. J. Fan, Y . Zhu, C. Liu, and G. Shi, “Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,” 2025

    2025

  9. [16]

    Bridging the sim-to-real gap for athletic loco-manipulation,

    N. Fey, G. B. Margolis, M. Peticco, and P. Agrawal, “Bridging the sim-to-real gap for athletic loco-manipulation,” 2025

    2025

  10. [18]

    Auto- matic LQR tuning based on Gaussian process global optimization,

    A. Marco, P. Hennig, J. Bohg, S. Schaal, and S. Trimpe, “Auto- matic LQR tuning based on Gaussian process global optimization,” inProceedings of IEEE International Conference on Robotics and Automation, (Stockholm, Sweden), pp. 270–277, 2016

    2016

  11. [19]

    Safe controller optimization for quadrotors with Gaussian processes,

    F. Berkenkamp, A. P. Schoellig, and A. Krause, “Safe controller optimization for quadrotors with Gaussian processes,” inProceed- ings of IEEE International Conference on Robotics and Automation, (Stockholm, Sweden), pp. 491–496, 2016

    2016

  12. [20]

    Bayesian gait optimization for bipedal locomotion,

    R. Calandra, N. Gopalan, A. Seyfarth, J. Peters, and M. P. Deisenroth, “Bayesian gait optimization for bipedal locomotion,” inProceedings of the International Conference on Learning and Intelligent Optimization, (Gainesville, FL, USA), pp. 274–290, 2014

    2014

  13. [21]

    Auto- matic gait optimization with Gaussian Process Regression,

    D. J. Lizotte, T. Wang, M. H. Bowling, and D. Schuurmans, “Auto- matic gait optimization with Gaussian Process Regression,” inPro- ceedings of the International Joint Conferences on Artificial Intelli- gence Organization, vol. 7, (Hyderabad, India), pp. 944–949, 2007

    2007

  14. [22]

    AutoTune: Controller tuning for high-speed flight,

    A. Loquercio, A. Saviolo, and D. Scaramuzza, “AutoTune: Controller tuning for high-speed flight,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 4432–4439, 2022

    2022

  15. [23]

    Policy search for model predictive control with application to agile drone flight,

    Y . Song and D. Scaramuzza, “Policy search for model predictive control with application to agile drone flight,”IEEE Transactions on Robotics, vol. 38, no. 4, pp. 2114–2130, 2022

    2022

  16. [24]

    Local policy search with Bayesian optimization,

    S. M ¨uller, A. von Rohr, and S. Trimpe, “Local policy search with Bayesian optimization,” inAdvances in Neural Information Processing Systems, 2021

    2021

  17. [25]

    A self- tuning LQR approach demonstrated on an inverted pendulum,

    S. Trimpe, A. Millane, S. Doessegger, and R. D’Andrea, “A self- tuning LQR approach demonstrated on an inverted pendulum,”IFAC Proceedings Volumes, vol. 47, no. 3, pp. 11281–11287, 2014

    2014

  18. [26]

    DiffLoop: Tuning PID controllers by differentiating through the feedback loop,

    A. R. Kumar and P. J. Ramadge, “DiffLoop: Tuning PID controllers by differentiating through the feedback loop,” inProceedings of the 55th Annual Conference on Information Sciences and Systems, (Baltimore, MD, USA), pp. 1–6, 2021

    2021

  19. [27]

    DiffTune: Autotuning through autodifferentiation,

    S. Cheng, M. Kim, L. Song, C. Yang, Y . Jin, S. Wang, and N. Ho- vakimyan, “DiffTune: Autotuning through autodifferentiation,”IEEE Transactions on Robotics, vol. 40, pp. 4085–4101, 2024

    2024

  20. [28]

    Diff- Tune+: Hyperparameter-free auto-tuning using auto-differentiation,

    S. Cheng, L. Song, M. Kim, S. Wang, and N. Hovakimyan, “Diff- Tune+: Hyperparameter-free auto-tuning using auto-differentiation,” in Learning for Dynamics and Control Conference, pp. 170–183, PMLR, 2023

    2023

  21. [29]

    DiffTune- MPC: Closed-loop learning for model predictive control,

    R. Tao, S. Cheng, X. Wang, S. Wang, and N. Hovakimyan, “DiffTune- MPC: Closed-loop learning for model predictive control,”IEEE Robotics and Automation Letters, 2024

    2024

  22. [30]

    High-dimensional Bayesian optimization using low-dimensional feature spaces,

    R. Moriconi, M. P. Deisenroth, and K. Sesh Kumar, “High-dimensional Bayesian optimization using low-dimensional feature spaces,”Ma- chine Learning, vol. 109, pp. 1925–1943, 2020

    1925

  23. [31]

    L 1Quad: L1 adaptive augmentation of geometric control for agile quadrotors with performance guarantees,

    Z. Wu*, S. Cheng*, P. Zhao, A. Gahlawat, K. A. Ackerman, A. Lakshmanan, C. Yang, J. Yu, and N. Hovakimyan, “L 1Quad: L1 adaptive augmentation of geometric control for agile quadrotors with performance guarantees,”IEEE Transactions on Control Systems Technology, 2025

    2025

  24. [32]

    JAX: composable transformations of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman- Milne, and Q. Zhang, “JAX: composable transformations of Python+NumPy programs,” 2018

    2018

  25. [33]

    Brax - a differentiable physics engine for large scale rigid body simulation,

    C. D. Freeman, E. Frey, A. Raichuk, S. Girgin, I. Mordatch, and O. Bachem, “Brax - a differentiable physics engine for large scale rigid body simulation,” 2021

    2021

  26. [34]

    Dojo: A differentiable simulator for robotics,

    A. H. Taylor, S. Le Cleac’h, Z. Kolter, M. Schwager, and Z. Manch- ester, “Dojo: A differentiable simulator for robotics,”arXiv preprint arXiv:2203.00806, 2022

    2022 arXiv

  27. [35]

    Neuralsim: Augmenting differentiable simulators with neural net- works,

    E. Heiden, D. Millard, E. Coumans, Y . Sheng, and G. S. Sukhatme, “Neuralsim: Augmenting differentiable simulators with neural net- works,” in2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 9474–9481, 2021

    2021

  28. [36]

    Differentiable simulation for physical system identification,

    Q. Le Lidec, I. Kalevatykh, I. Laptev, C. Schmid, and J. Carpentier, “Differentiable simulation for physical system identification,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3413–3420, 2021

    2021

  29. [37]

    gradsim: Differentiable simulation for system identification and visuomotor control,

    J. K. Murthy, M. Macklin, F. Golemo, V . V oleti, L. Petrini, M. Weiss, B. Considine, J. Parent-L ´evesque, K. Xie, K. Erleben, L. Paull, F. Shkurti, D. Nowrouzezahrai, and S. Fidler, “gradsim: Differentiable simulation for system identification and visuomotor control,” inInte...

    2021

  30. [38]

    Learning quadruped locomotion using differentiable simulation,

    Y . Song, S. bae Kim, and D. Scaramuzza, “Learning quadruped locomotion using differentiable simulation,” in8th Annual Conference on Robot Learning, 2024

    2024

  31. [39]

    Residual policy learning for perceptive quadruped control using differentiable simulation,

    J. Y . Luo, Y . Song, V . Klemm, F. Shi, D. Scaramuzza, and M. Hutter, “Residual policy learning for perceptive quadruped control using differentiable simulation,” 2024

    2024

  32. [40]

    Training efficient controllers via analytic policy gradient,

    N. Wiedemann, V . W ¨uest, A. Loquercio, M. M ¨uller, D. Floreano, and D. Scaramuzza, “Training efficient controllers via analytic policy gradient,” in2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 1349–1356, 2023

    2023

  33. [41]

    Some notes on alternating opti- mization,

    J. C. Bezdek and R. J. Hathaway, “Some notes on alternating opti- mization,” inAdvances in Soft Computing, 2002

    2002

  34. [42]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” 2017

    2017

  35. [43]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033, IEEE, 2012

    2012

  36. [44]

    Flax: A neural network library and ecosystem for JAX,

    J. Heek, A. Levskaya, A. Oliver, M. Ritter, B. Rondepierre, A. Steiner, and M. van Zee, “Flax: A neural network library and ecosystem for JAX,” 2024

    2024

  37. [45]

    The DeepMind JAX Ecosystem,

    G. DeepMind, “The DeepMind JAX Ecosystem,” 2020

    2020

  38. [46]

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

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

    2023

  39. [47]

    Physx sdk

    NVIDIA, “Physx sdk.”

  40. [48]

    Barkour: Benchmarking animal-level agility with quadruped robots,

    K. C. et al., “Barkour: Benchmarking animal-level agility with quadruped robots,” 2023

    2023

  41. [49]

    Pybullet, a python module for physics simulation for games, robotics and machine learning

    E. Coumans and Y . Bai, “Pybullet, a python module for physics simulation for games, robotics and machine learning.”http:// pybullet.org, 2016–2021

    2016

  42. [50]

    Force-and-moment-based model predictive control for achieving highly dynamic locomotion on bipedal robots,

    J. Li and Q. Nguyen, “Force-and-moment-based model predictive control for achieving highly dynamic locomotion on bipedal robots,” in2021 60th IEEE Conference on Decision and Control (CDC), pp. 1024–1030, 2021

    2021

  43. [51]

    Efficient and modular implicit differentiation,

    M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares- L´opez, F. Pedregosa, and J.-P. Vert, “Efficient and modular implicit differentiation,”arXiv preprint arXiv:2105.15183, 2021

    2021 arXiv

  44. [52]

    Dynamic loco- manipulation on hector: Humanoid for enhanced control and open- source research,

    J. Li, J. Ma, O. Kolt, M. Shah, and Q. Nguyen, “Dynamic loco- manipulation on hector: Humanoid for enhanced control and open- source research,” 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.