acceptodds
Withdrawn from ICLR 2027

EPIC-PIML: Epistemic-Constructed Physics-Informed Machine Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning the dynamics of complex robotic systems via a hybrid physics-informed machine learning (HPIML) approach, which combines known physics with machine learning (ML), has shown promising performance in decision-making and optimal control tasks. However, under environmental uncertainty, HPIML is not necessarily a robust predictive model: the in-distribution region favors the ML term, while the out-of-distribution region favors the physics term. Therefore, adaptively weighing the outputs of these two terms across different environments or regions is important for an accurate and robust learning model. This paper proposes Epistemic-Constructed Physics-Informed Machine Learning (EPIC-PIML), which introduces an uncertainty mechanism that coordinates the outputs of physics and ML to produce a well-adjusted prediction from the HPIML model. Specifically, we attach an Epistemic Neural Network to a pre-trained HPIML model to estimate its epistemic uncertainty in closed form, and this uncertainty then drives the mechanism that determines the relative contribution of the physics and ML terms to the HPIML model's overall prediction. We prove that this mechanism converges to the correct trade-off as data and capacity grow, with a matching guarantee that its uncertainty estimate yields valid prediction sets even out-of-distribution. We further embed EPIC-PIML in a model predictive control framework, enabling closed-loop robotic control. We demonstrate our proposed method across two robotic systems at different scales: a quadrotor and an autonomous racing car. Across these experiments, the proposed model outperforms four baselines in system modeling and control performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Physics-informed machine learning, epistimic uncertainty, model predictive control, robotics

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_5bt3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper develops an uncertainty-aware hybrid physics-informed machine learning architecture that attaches an Epistemic Neural Network to a pre-trained HPIML model. Based on this, the paper also proposes an adaptive mechanism, built from monotone function approximators, that uses this uncertainty to rebalance the physics and ML terms of HPIML, with theoretical guarantees on both its convergence and the coverage of its uncertainty estimate.

Critical strengths

This paper develops an uncertainty-aware hybrid physics-informed machine learning architecture that attaches an Epistemic Neural Network to a pre-trained HPIML model.

Critical weaknesses
  1. I do not think the EPIC-PIML-Based MPC is a novel contribution. It is more like a direct use of the learned dynamics model for MPC. Besides, I suggest a critical comparison of EPIC-PIML-based MPC with an adaptive controller, rather than the simple PHY-based MPC. At a high level, they are similar. The adaptive controller uses the real-time states to estimate model mismatch or uncertainty, and then uses the estimation to compute a control input to compensate for the unknown model mismatch. More importantly, adaptive control does not face the latency issue.
  2. The comparison experiments are not sufficient and lack critical information. Specifically, the paper compares PIML. In recent years, significant frameworks for PIML have been proposed, such as deep Koopman, Physical World Model, and Physics-Enhanced State Space Model. Which models are used for comparison? Moreover, the critical information about the training model is missing. For example, how are the control commands in the training dataset obtained? Are the control commands in the training dataset realizable? If they are reliable, why not directly train the network to output the control commands (given input of system states) and compare it to the MPC? Meanwhile, the paper claims that the proposed framework can address the out-of-distribution problem well. To have a more convincing demonstration, the testing dataset should have a different distribution from the training and evaluation datasets, rather than only from different tracks.
Questions for authors
  1. Some critical descriptions are missing, rendering the paper hard to follow. For example, in Eqs. (2) and (3), what are the meanings of z_{k} (outputs of the previous NN)? What is the physical meaning of the reference features? I suggest adding concrete examples to clarify these points.
  2. Section 4.2 discusses the adaptive weight \alpha but does not present the detailed functional forms. Are the functions unique?
  3. I do not think the EPIC-PIML-Based MPC is a novel contribution. It is more like a direct use of the learned dynamics model for MPC.
  4. The comparison experiments are not sufficient and lack critical information. For example, the paper compares PIML. In recent years, significant frameworks for PIML have been proposed, such as deep Koopman, Physical World Model, and Physics-Enhanced State Space Model. Which models are used for comparison?
  5. The critical information about the training model is missing. For example, how are the control commands in the training dataset obtained? Are the control commands in the training dataset realizable? If they are reliable, why not directly train the network to output the control commands (given input of system states) and compare it to the MPC?
  6. I suggest a critical comparison of EPIC-PIML-based MPC with an adaptive controller, rather than the simple PHY-based MPC. Because, at a high level, they are similar. The adaptive controller uses the real-time states to estimate model mismatch or uncertainty, and then uses the estimation to compute a control input to compensate for the unknown model mismatch. More importantly, adaptive control does not face the latency issue.
  7. All the experiments are performed in simple simulators, implying that the obtained theoretical results may not hold in the real physical world, as the physical operating environments are much more complex and non-stationary.
  8. The paper claims that the proposed framework can address the out-of-distribution problem well. To have a more convincing demonstration, the testing dataset should have a different distribution from the training and evaluation datasets, rather than only from different tracks.
Reviewer_2tvmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EPIC-PIML, an uncertainty-aware hybrid dynamics learning framework that combines a nominal physics model with a learned residual model. It uses an epistemic neural network to estimate predictive uncertainty and a monotonic gating mechanism to adjust the contribution of the learned residual. The model is integrated into MPPI control and evaluated on quadrotor and autonomous racing simulations, demonstrating improved prediction accuracy and tracking performance.

Critical strengths

The paper addresses an important problem in learning-based control. The uncertainty-dependent gating mechanism is intuitive and well-structured, with theoretical analysis under stated assumptions. Experiments on two robotic systems demonstrate promising improvements in long-horizon prediction accuracy and closed-loop tracking performance.

Critical weaknesses
  1. The methodological novelty needs further clarification. Neural residual dynamics learning and adaptive model-based control have been extensively studied, including Jiahao et al. (CoRL 2023), Sun et al. (RA-L 2021), and Smith and Mistry (RA-L 2020). The proposed approach combines existing uncertainty estimation and monotonic gating techniques, but the experiments only compare against conventional physics-based and learning-based models. Comparisons with uncertainty-aware residual models, online adaptive methods, and simpler gating strategies are needed to establish its advantages.

  2. The theoretical guarantees seem to rely on strong assumptions. Theorem 1 assumes the true residual can be recovered through a monotone function of estimated uncertainty, without much justification. Theorem 2 relies on a Gaussian posterior assumption and does not establish that the estimated covariance accurately captures prediction errors, particularly out of distribution.

  3. The experiments primarily consider changes in operating conditions and injected Gaussian disturbances under fixed dynamics. Evaluation under changing system parameters or persistent disturbances would better demonstrate its practical applicability.

Jiahao, Tom Z., Kong Yao Chee, and M. Ani Hsieh. "Online dynamics learning for predictive control with an application to aerial robots." Conference on Robot Learning. PMLR, 2023.

Sun, Yu, et al. "Online learning of unknown dynamics for model-based controllers in legged locomotion." IEEE Robotics and Automation Letters 6.4 (2021): 8442-8449.

Smith, Joshua, and Michael Mistry. "Online simultaneous semi-parametric dynamics model learning." IEEE Robotics and Automation Letters 5.2 (2020): 2039-2046.

Questions for authors
  1. How does the proposed method compare with existing uncertainty-aware residual models and online adaptive methods?

  2. Can the proposed method meet real-time control requirements given its computational cost?

Reviewer_qm8ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EPIC-PIML, an uncertainty-aware hybrid physics-informed machine-learning model for robotic dynamics and control. Starting from a pretrained model of the form f_phy(z)+f_ml(z), the authors attach a linear Epistemic Neural Network (ENN) whose epistemic covariance can be computed in closed form. A monotone decreasing gate then scales the learned residual: the model trusts the ML correction when epistemic uncertainty is low and falls back toward the physics model when uncertainty is high. The paper provides a min-max monotone gate and a Kalman-style gate, a staged training procedure, and an MPPI-based controller using EPIC-PIML rollouts. Theoretical results give a convergence bound under a realizable monotone-gate assumption and a conditional prediction-set coverage result under a Gaussian linear-ENN interpretation. Experiments on a simulated quadrotor and an F1TENTH racing car show improved long-horizon position RMSE and selected closed-loop control metrics over fixed HPIML and other baselines, at the cost of substantially higher computation.

Critical strengths
  1. The method is simple, well motivated. The paper addresses an important failure mode of hybrid physics–ML models under distribution shift.
  2. The monotone gate gives an interpretable inductive bias, while the linear ENN provides a computationally convenient closed-form covariance for online planning.
  3. The evaluation includes both long-horizon open-loop modeling and closed-loop MPPI control on two robotic systems.
Critical weaknesses
  1. The theoretical guarantees rely on strong assumptions. In particular, Assumption 1 directly assumes that the true residual can be represented by a coordinate-wise monotone function of the estimated uncertainty. Theorem 1 therefore does not establish general OOD convergence, especially outside the support of the training distribution.
  2. The component contributions and baseline fairness are not fully isolated. In particular, the paper lacks ablations for the ENN, monotone gate, final refitting, and alternative uncertainty estimators. Moreover, the risk coefficient is tuned for EPIC-PIML but set to zero for the baselines in the quadrotor experiment.
  3. The reported runtime of 119.4 ms per step for EPIC-PIML versus 33.2 ms for HPIML raises concerns for the claimed 50 Hz control setting. The paper should report whether the method meets the real-time budget and how performance changes under reduced MPPI computation.
Questions for authors
  1. Can the authors report empirical uncertainty calibration: coverage versus nominal level, prediction-set size, and the correlation between Sigma_ii and one-/multi-step prediction error?
  2. What evidence supports the realizable monotone-gate assumption in both environments? How does performance change when the residual cannot be represented as a function of Sigma_ii alone?
  3. Can the authors provide multi-seed results and ablations for the ENN, monotone gate, physics fallback, risk penalty, prior scale, and gate family?
  4. How is the non-centrality parameter used in Eq. (15) obtained at deployment when the true next state is unknown?
  5. How should the claimed OOD prediction-set validity be interpreted given the strong assumptions of Theorem 2?
Reviewer_feDmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EPIC-PIML, a physics-informed hybrid dynamics model for predictive control. It uses an Epistemic Neural Network (ENN) to estimate uncertainty, then applies a monotonic gate to regulate the learned dynamics residual. The authors evaluate open-loop prediction and closed-loop control on a quadrotor and an F1TENTH racing car. The paper also claims that its uncertainty estimates yield valid out-of-distribution prediction sets. Presentation: 3 — Clear and easy to follow The research question and model components are presented clearly. The relationship among the uncertainty estimate, residual gate, and predictive controller is understandable, although several claims need more careful qualification. Informativeness: 2 — I learned something interesting or useful The combination of epistemic uncertainty, a gated residual model, and predictive control is a useful design direction for robotics. Soundness: 1 — Significant concerns The prediction-set guarantee rests on a strong error-distribution assumption, and the control experiments do not fully isolate the effect of the proposed model and gate. The reported computation time also leaves the real-time control claim unresolved.

Critical strengths
  1. The paper addresses a relevant problem: using learned dynamics while accounting for uncertainty in model-based robot control.
  2. Equations (5)–(8) provide a concrete and implementable model design. The evaluation includes both prediction and downstream control, as well as two distinct robotic systems.
  3. The reported racing-car result is promising. In Fig. 4, EPIC-PIML achieves a 20-step position RMSE of 7.956 cm, lower than the four listed baselines.
Critical weaknesses

The prediction-set claim exceeds the demonstrated guarantee. The proof of Theorem 2 in Appendix C.2 treats the true next-state error as Gaussian with the covariance produced by the ENN. Variation across ENN outputs does not, by itself, establish that real-system prediction errors follow that distribution, particularly under distribution shift. The theorem therefore supports conditional one-step coverage under its stated model assumption; it does not establish empirical out-of-distribution coverage. 2. The closed-loop comparisons do not isolate the contribution of the proposed model. The experiments mainly compare predictive models within a chosen planning framework. They do not establish how the resulting controller compares with other practical control approaches on both tracking accuracy and computation cost. 3. The reported timing does not support 50 Hz real-time operation. The racing-car control period is 20 ms, while the reported average computation time for EPIC-PIML is 119.4 ms per step (lines 466–471). The paper needs to explain how planning time and action delay are handled in its control evaluation. 4. The quadrotor result confounds the model and planning objective. Equation (17) includes an uncertainty-based risk penalty, while Appendix D.1 states that the risk weight is zero for the baselines and tuned for EPIC-PIML. The comparison in Fig. 3 therefore changes both the predictive model and the control objective. The current evidence cannot attribute the full closed-loop improvement to the proposed gate.

Questions for authors

What evidence shows that the ENN covariance is calibrated to real next-state prediction errors, especially on independent out-of-distribution data? Please report empirical coverage alongside prediction-set size. 2. Can the authors evaluate EPIC-PIML and the baselines under the same planning objective, then report EPIC-PIML with and without the uncertainty-based risk penalty? 3. Does the simulator wait for each planning computation before advancing? How are action delays represented? What control performance is obtained under a strict 20 ms computation budget? 4. How does the proposed controller compare with suitable practical alternatives, such as conventional MPC or task-appropriate classical controllers, when control accuracy and computation cost are measured together? 5. Which statements about uncertainty and robustness are assumptions or modeling choices, and which are directly supported by experiments?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.