Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Learning Causal Structure Distributions for Robust Planning

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper claims that a robot's learned dynamics model is more robust to sensor failure and dynamics changes when it samples causal masks from a distribution over structures rather than relying on a single causal graph.

desk verdict The method is a cheap, genuinely robust latent-masking trick, but the causal mechanism is unproven — the paper deserves review for the empirics, not the causal framing. read the letter →

arxiv 2508.06742 v1 pith:W7JZQQK3 submitted 2025-08-08 cs.RO cs.AIcs.LGcs.SYeess.SY

classification cs.ROcs.AIcs.LGcs.SYeess.SY
keywords causalstructuredistributionintegratedgradientsdynamicslearningmodel-basedplanningrobustnessencoder-multidecoderstructuralmodellatentmasking
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 learned dynamics model becomes more robust to real-world disruptions if, instead of committing to a single causal graph, it samples masks from a distribution over causal structures. The authors build CADY, an encoder-multidecoder network whose latent features are masked by binary matrices drawn from a Bernoulli distribution; the edge probabilities are estimated with Integrated Gradients on a fully-wired network. Across Cartpole, Pusher, and a simulated and real Jackal ground vehicle, CADY keeps planning accurate when sensors freeze or go noisy, adapts quickly to unmodeled control-gain interventions, and uses 20x-250x fewer FLOPs and parameters than the PETS ensemble baseline. The broader claim is that accounting for structural uncertainty is a cheap way to make model-based planning robust.

What carries the argument

The load-bearing object is the causal structure distribution $P(p)$, a product of independent Bernoulli variables $E_{ij}$ indicating an edge from input $i$ to next-state variable $j$. Its parameters are set to normalized Integrated Gradients, $p_{ij} := \mathrm{IG}_{ij}(x, x'=0; f_C)$, clipped to $[\rho_{\min}, 1-\rho_{\min}]$, so feature attribution stands in for causal discovery. Every forward call samples a mask $M \sim P(p)$ and applies it in the latent space, $\tilde{z}_j = z \odot m_{:,j}$, before a per-output decoder produces a Gaussian over the next state. This turns structural uncertainty into a cheap training-time regularizer and an inference-time marginalization.

What would settle it

On a system with a known causal graph, compare CADY's behavior under three mask sources: Integrated-Gradient estimates, the ground-truth graph, and uniform random masks with matched sparsity. If random masks match IG masks in robustness, the causal estimator is not carrying the result; if ground-truth masks do not beat IG masks, the structure distribution itself is not the cause. A second check: rescale one input feature by a constant and see whether p_ij shifts; IG is sensitive to scaling, so a causally meaningful estimator should not arbitrarily flip the masks.

Watch

Extended reading notes

Core claim

The discovery is that uncertainty over causal structure, not just over function parameters, belongs inside a learned dynamics model. CADY represents the dynamics as $f(x) = f_{\mathrm{dec}}(f_{\mathrm{enc}}(x) \odot M)$ with $M \sim P(p)$, where $P$ is a product of independent Bernoulli edge distributions over the bipartite graphs allowed by the Markovian state-transition assumption. Each $p_{ij}$ is estimated as a normalized, clipped Integrated Gradient of a fully-wired contribution model, and every forward pass draws a fresh causal mask. In experiments this marginalization over plausible structures yields up to 13% lower performance degradation under frozen or noisy sensors, one-step post-

Load-bearing premise

The load-bearing premise is that Integrated Gradients on a fully-wired network trained on the same observational data yield valid edge probabilities for a causal structure distribution; if attribution values do not track true causal effects, the masks can drop predictive features or keep spurious ones and the claimed mechanism collapses.

Editorial extensions

If this is right

  • A dynamics model that samples causal masks at inference time degrades 1.6-13% less than PETS when a sensor freezes or becomes noisy, with the gap growing in higher-dimensional systems.
  • Under an unmodeled control-gain intervention, CADY's one-step prediction error is 0.124 versus 2.039 for PETS, and fine-tuning on recent data recovers faster.
  • On the Jackal ground vehicle under GPS-like state noise (variance 0.01), mission success rises to 85.7% from PETS's 55.7%; on a steep mountain mission, CADY keeps 27% success where PETS falls to 2.8%.
  • Marginalizing over a distribution of structures rather than fixing one graph improves performance by 3.2% on Cartpole and 5.8% on Pusher.
  • CADY uses 20x fewer FLOPs and parameters than PETS on Pusher and more than 250x fewer on Jackal, including the cost of sampling the mask.

Reading between the lines

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

  • If the mechanism is general, the same encoder-multidecoder plus Bernoulli-mask recipe could apply to any controlled dynamical system with sparse interactions, not just robots.
  • A natural testable extension is to replace the Integrated-Gradients estimator with a proper Bayesian structure-learning method; if the robustness gains survive that swap, the architecture is the contribution and the attribution estimator is incidental.
  • The authors' own Cartpole sanity check shows one mismatched edge ($\dot{\theta}_t \to \dot{x}_{t+1}$); intervening on that specific edge would be a sharper test of whether the estimated masks track true causal structure.
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

3 major / 6 minor

Summary. The paper proposes CADY, a probabilistic encoder-multidecoder dynamics model that masks its latent features with binary matrices sampled from a learned distribution over causal structures. The structure distribution is estimated by computing Integrated Gradients on a separately trained 'fully wired' network and normalizing the attributions to Bernoulli probabilities (Eqs. 3–4). The authors argue that marginalizing over structural uncertainty yields dynamics models that are more robust to missing/noisy observations and unexpected interventions, while using far fewer parameters and FLOPs than PETS. Experiments cover Cartpole, Pusher, and a Jackal ground vehicle in simulation and the real world, with robustness, intervention, and computational-cost evaluations.

Significance. If the causal-structure claim is substantiated, the paper offers a computationally cheap alternative to causal discovery for improving model-based planning robustness. The empirical evaluation is genuinely broad and includes several well-designed controls: held-out noise and missing-feature tests, a parameter-count control PETS(S) in the intervention study, and a real-world zero-shot deployment. The reported gains are large (e.g., Jackal Mission 1 success 85.7% vs. 55.7% at noise variance 0.01; post-intervention one-step MSE 0.124 vs. 2.039) and the computational savings are substantial. However, the central mechanism—that the Integrated-Gradients-derived mask probabilities carry causal information—is not isolated from generic stochastic masking or feature selection, so the paper's headline causal explanation remains conditional on additional controls.

major comments (3)
  1. [Section IV-A, Eqs. (3)–(4)] The causal structure distribution p_ij is defined as normalized Integrated Gradients of a network f_C trained on the same observational dataset D used to train the dynamics model. As the paper itself states, 'IG was not designed as a causality test.' The only ground-truth validation is the Cartpole comparison in Fig. 6, which visibly underestimates the θ̇→ẋ edge and is in the same environment used for training. The central claim that masks sampled from P(p) are 'causally informed' is thus not yet established. A concrete test: compare CADY's performance against the same architecture with (i) uniform p=0.5, (ii) p with shuffled entries, and (iii) IG computed from a model trained on label-randomized data. If these controls perform comparably, the gains are due to stochastic masking/regularization, not causal structure. This is load-bearing for the abstract's causal claim and must be address
  2. [Section V-D, fixed-graph ablation] The ablation showing that resampling from the learned p improves performance over a fixed graph is presented as evidence for the value of structural uncertainty. However, the fixed graph is presumably derived from p itself, and the comparison conflates the specific learned probabilities with the effect of stochastic mask sampling. Without a uniform-p or shuffled-p control, the result cannot distinguish causal information from generic mask diversity. The paper should report the fixed-graph result alongside the proposed uniform/shuffled controls, and clarify how the 'fixed graph' is selected (threshold, most-likely edge, etc.).
  3. [Section V-B, Fig. 6] The structural validation is limited to a single environment (Cartpole) and even there the estimated distribution misses the true θ̇→ẋ edge with low probability. For Pusher and Jackal no ground-truth structure is provided at all, yet the robustness narrative depends on the masks being causally meaningful in those settings. If the causal claim is retained, the authors should either provide structural validation in at least one more environment or show that the missing edge is not load-bearing for the robustness results. Alternatively, the text should be revised to characterize the masks as 'predictive-feature masks' rather than causal structures, with the causal terminology softened accordingly.
minor comments (6)
  1. [Section IV-A, Eq. (4)] The smoothing function s(·) and the clipping bound ρ_min are introduced in Eq. (4) but only defined in Section V-A ('cubic root' and 0.02). Define them at first use for clarity.
  2. [Section V-A, Baselines] PETS(S) is first mentioned in Fig. 10 but not defined in the baselines list; please define it explicitly (presumably 'PETS with reduced parameters') at first occurrence.
  3. [Fig. 6] The figure caption says the ground truth is for the state delta, but the text 'the figure does not show the ∆ symbol for ease of visualization' appears only in the body; please state this in the caption to prevent misinterpretation of the x_t → x_{t+1} relation.
  4. [References] Reference [25] appears in the bibliography between [24] and [26] but is not in numeric order in the text citation sequence; please reorder or renumber.
  5. [Table I] The parameter counts for CADY are very small (e.g., 230 for Cartpole). It would help to report the architecture breakdown (encoder/decoder sizes) or at least confirm that these counts include all trainable parameters.
  6. [Throughout] Minor typos: 'FLOPS' vs. 'FLOPs' is inconsistent; 'P(pD)' in Section IV-B is used before defining p_D. Please proofread notation consistency.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the IG-based structure estimate is a fixed modeling input, and the robustness results are measured on held-out conditions rather than re-derived from the fitted p.

full rationale

The paper's central derivation chain is not circular. The causal structure distribution P(p) is computed by a fixed formula (Eqs. 3-4) from the Integrated Gradients of a separately trained contribution model f_C on dataset D, and then used to sample masks for the dynamics model f_D. This is a modeling choice, not a reduction of the output to the input: the claimed robustness gains are measured on corrupted inputs, interventions, and held-out post-intervention conditions, and these measurements are not used to define p. The paper explicitly acknowledges a validity caveat in Section IV-A: 'Although IG was not designed as a causality test, we use it to estimate the likelihood of each feature being a cause in the underlying SCM.' That caveat concerns whether the attributions track true causal effects, which is a correctness/validity concern rather than a circularity concern. Similarly, the acknowledged mismatch in Fig. 6 (the theta-dot to x-dot edge) is an empirical validation failure, not a logical circularity. The only self-citation is reference [37] for the general statement that observational data can exhibit bias, which is not load-bearing for any core claim. A missing control (e.g., uniform p=0.5 or shuffled p) would be an ablation/comparison weakness, but it does not make the derivation circular. Therefore the paper is self-contained with respect to the circularity-specific failure modes considered here.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the assumed equivalence between feature-attribution scores computed on a fitted network and probabilities of causal edges; that is the single largest unproven input. Secondary inputs are the bipartite/Markovian restriction, edge independence, and the validity of the zero null input for Integrated Gradients, none of which is empirically checked. Tuned constants with stated empirical selection are rho_min and the cubic-root smoothing; planner hyperparameters are standard defaults. No new physical entities are introduced; the causal-structure distribution is a modeling construct whose only external check reuses the training environment.

free parameters (6)
  • rho_min = 0.02
    Clipping bound on Bernoulli parameters in Eq. (4); the authors state it 'empirically found improved performance', i.e., selected on the evaluation tasks.
  • smoothing function s(.) = cubic root
    Eq. (4) smoothing chosen by hand to prevent one attribution from dominating; no principled justification.
  • causal structure parameters p_D = estimated via IG, initialized at 0.5
    The structure distribution itself: derived from f_C trained on the same dataset D, normalized per column and clipped; not obtained from causal discovery or interventions.
  • batch size = 8
    Chosen to 'enforce frequent causal graph re-sampling' during training; controls mask stochasticity and hence the effective marginalization over structures.
  • decoder hidden sizes = 3, 13, 20 (Cartpole, Pusher, Jackal)
    Per-environment capacity choices; much smaller than PETS by design, which is what drives the reported FLOPs/parameter reductions.
  • planner hyperparameters (CEM alpha, elite ratio; MPPI gamma, sigma, beta) = 0.1, 0.1; 0.9, 0.01, 0.6
    Sampling-based planner settings that directly affect mission success and time in Table II and Fig. 9; standard defaults but not justified per environment.
assumptions (6)
  • ad hoc to paper Integrated Gradients attributions of a fully-wired network f_C trained on observational data D approximate the probabilities of true causal edges in the robot's SCM (Section IV-A: p_ij := IG_ij(x, x'; M), Eq. 3).
    Load-bearing, unvalidated: the authors state IG 'was not designed as a causality test' and offer only a heuristic justification ('implicitly assigning blame to a certain cause'). No intervention data, counterfactual check, or theory supports the causal reading; the Cartpole sanity check (Fig. 6) is on the same environment that produced the training data.
  • domain assumption The dynamics causal graph is bipartite and Markovian: edges exist only from (s_t, a_t) to s_{t+1} (Section IV-A, 'Bounding the SCM Search Space').
    Reduces the DAG search space to (n+p) by n edge indicators. Valid for the studied first-order platforms, but excludes hidden state, time-delayed effects, and within-timestep variable interactions.
  • domain assumption Edge indicators E_ij are independent Bernoulli variables; the joint PMF factorizes (Eq. 1-2).
    Treats structural uncertainty as independent per edge, ignoring correlations among candidate parents, e.g., two actuators jointly affecting one joint.
  • domain assumption The null input x' = 0 satisfies M(x') ≈ 0 and is therefore a valid no-cause baseline for Integrated Gradients (Section III-B, used in Eq. 3).
    The paper learns state deltas so zero input plausibly yields zero output, but this holds only when variables are normalized around zero; no verification is reported, and for angular variables the zero baseline is arbitrary.
  • standard math Each decoder outputs a Gaussian over next-state deltas, capturing aleatoric uncertainty (Section IV-B).
    Standard probabilistic-dynamics modeling assumption inherited from PETS [41]; not new but unverified for the real Jackal.
  • domain assumption The structure distribution estimated during training (from D) remains valid at deployment, in new environments, and under dynamics changes (Section IV-B inference description; Abstract claim of 'new tasks in novel environments').
    p_D is frozen at inference; the intervention experiments keep the original masks and only fine-tune weights, so adaptation relies on the pre-intervention structure distribution being approximately correct.
invented entities (1)
  • Causal structure distribution P(p) over edge indicators E_ij
    purpose: Models epistemic uncertainty over the causal graph of the robot's dynamics; sampled masks select which latent features feed each next-state decoder.
    A modeling construct computed from attributions of a network trained on the same observational dataset as the dynamics model. Its only external check is the Cartpole ground-truth comparison (Fig. 6), a sanity check on the same data-generating environment, not an independent falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Causal Structure Distributions for Robust Planning." pith.science (2026). https://pith.science/paper/W7JZQQK3

@misc{pith2026250806742,
  author       = {Pith},
  title        = {Pith review of: Learning Causal Structure Distributions for Robust Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7JZQQK3}},
  note         = {Machine review of arXiv:2508.06742}
}
read the original abstract

Structural causal models describe how the components of a robotic system interact. They provide both structural and functional information about the relationships that are present in the system. The structural information outlines the variables among which there is interaction. The functional information describes how such interactions work, via equations or learned models. In this paper we find that learning the functional relationships while accounting for the uncertainty about the structural information leads to more robust dynamics models which improves downstream planning, while using significantly lower computational resources. This in contrast with common model-learning methods that ignore the causal structure and fail to leverage the sparsity of interactions in robotic systems. We achieve this by estimating a causal structure distribution that is used to sample causal graphs that inform the latent-space representations in an encoder-multidecoder probabilistic model. We show that our model can be used to learn the dynamics of a robot, which together with a sampling-based planner can be used to perform new tasks in novel environments, provided an objective function for the new requirement is available. We validate our method using manipulators and mobile robots in both simulation and the real-world. Additionally, we validate the learned dynamics' adaptability and increased robustness to corrupted inputs and changes in the environment, which is highly desirable in challenging real-world robotics scenarios. Video: https://youtu.be/X6k5t7OOnNc.

Figures

Figures reproduced from arXiv: 2508.06742 by the authors.

Figure 1
Figure 1. Proposed CAusally-informed DYnamics (CADY) model. We train a contribution model f C using the robot’s experience to estimate a distribution over causal structures P(p). A probabilistic encoder-multidecoder f D (inside the dotted box) masks its latent space using sampled causal structures from P(p), leading to a dynamics model that improves downstream planning robustness. However, in most cases, the true SCM is unkno… view at source ↗
Figure 2
Figure 2. Dynamics causal graph with n = 3 state ([xt, yt, θt]) and p = 2 con￾trol ([vt, ωt]) variables for a discretized differential drive robot model [40]. For clarity, functional relationships are shown in the dotted box, and the discretization constant ∆t is omitted. Modelling the SCM Distribution. We propose to learn the distribution over causal structures of the robot’s dynamics by modelling the distribution P of the e… view at source ↗
Figure 4
Figure 4. Our model learns faster than other causal, model-based and model-free baselines trained in the Cartpole environment. Statistics (mean ± std) computed over five runs. 0 5000 10000 15000 20000 25000 30000 Timestep −300 −250 −200 −150 −100 Reward Causal methods (Pusher) Ours REG CMI 0 5000 10000 15000 20000 25000 30000 Timestep −300 −250 −200 −150 −100 −50 Reward Non-causal methods (Pusher) Ours PETS SAC TD3 PPO [PITH… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Using lower computational resources, our model achieves similar performance as other causal, model-based and model-free baselines trained in the Pusher environment. Statistics (mean ± std) computed over five runs (three runs for CMI and Reg). B. Learned Models We begin…
Figure 6
Figure 6. Figure 6: Estimated Causal Structure Distribution P(p) for the Cartpole model. (Left) Ground truth causal graph. (Right) Estimated distribution P(p). Each cell’s color represents the probability p that an edge exists between two nodes. We can also appreciate the bounded (n + p) …
Figure 7
Figure 7. Figure 7: Performance as a function of computational requirements. Note the broken x-axis for ease of visualization. FLOPs (computed with [50]) and parameter count, with improvements consistently exceeding one to two orders of magnitude. Note that these results consider the over…
Figure 8
Figure 8. Figure 8: Cartpole and Pusher Robustness Evaluation. Performance degradation (%) when the inputs have missing or noisy features. 0.0 0.2 0.4 0.6 0.8 1.0 Noise Variance −0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Mean ± std ↑ Success Rate Ours CMI PETS 0.0 0.2 0.4 0.6 0.8 1.0 Noise Variance…
Figure 9
Figure 9. Figure 9: Jackal navigation performance with noisy state estimates. Each row shows the results for different missions. Mean ± std. dev. over 10 trials. Only successful runs are considered for the total time and traveled distance statistics. D. Increased Robustness to Missing and…
Figure 10
Figure 10. Figure 10: Effect of Unexpected Interventions and Online Corrections. Mean Squared Error for one-step prediction with each dynamics model before the intervention (pre-interv.), after the intervention (post-interv.), and after fine-tuning while remaining in the intervened dynamic…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 44 canonical work pages

  1. [1]

    Causality

    Judea Pearl. Causality. Cambridge university press, 2009

  2. [2]

    Toward causal representation learning

    Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE , 2021

  3. [3]

    The cross-entropy method for optimization

    Zdravko I Botev, Dirk P Kroese, Reuven Y Rubinstein, and Pierre L’Ecuyer. The cross-entropy method for optimization. In Handbook of statistics, volume 31, pages 35–59. Elsevier, 2013

  4. [4]

    Robust model predictive path integral control: Analysis and performance guarantees

    Manan S Gandhi, Bogdan Vlahov, Jason Gibson, Grady Williams, and Evangelos A Theodorou. Robust model predictive path integral control: Analysis and performance guarantees. IEEE RA-L, 2021

  5. [5]

    Counting unlabeled acyclic digraphs

    Robert W Robinson. Counting unlabeled acyclic digraphs. In Combina- torial Mathematics V: Proceedings of the Fifth Australian Conference, Held at the Royal Melbourne Institute of Technology, August 24–26, 1976, pages 28–43. Springer, 1977

  6. [6]

    Car-despot: Causally-informed online pomdp planning for robots in confounded environments

    Ricardo Cannizzaro and Lars Kunze. Car-despot: Causally-informed online pomdp planning for robots in confounded environments. Inter- national Conference on Intelligent Robots and Systems (IROS) , 2023

  7. [7]

    Causalaf: Causal autoregressive flow for safety-critical driving scenario generation

    Wenhao Ding, Haohong Lin, Bo Li, and Ding Zhao. Causalaf: Causal autoregressive flow for safety-critical driving scenario generation. In Conference on Robot Learning , pages 812–823. PMLR, 2023

  8. [8]

    Causal dynamics learning for task-independent state abstraction

    Zizhao Wang, Xuesu Xiao, Zifan Xu, Yuke Zhu, and Peter Stone. Causal dynamics learning for task-independent state abstraction. arXiv preprint arXiv:2206.13452, 2022

Show all 50 references
  1. [9]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning , pages 3319–3328. PMLR, 2017

  2. [10]

    Causal diagrams for empirical research

    Judea Pearl. Causal diagrams for empirical research. Biometrik, 1995

  3. [11]

    Adarl: What, where, and how to adapt in transfer reinforcement learning

    Biwei Huang, Fan Feng, Chaochao Lu, Sara Magliacane, and Kun Zhang. Adarl: What, where, and how to adapt in transfer reinforcement learning. In ICLR, 2022

  4. [12]

    Learning dynamic attribute-factored world models for efficient multi-object reinforcement learning

    Fan Feng and Sara Magliacane. Learning dynamic attribute-factored world models for efficient multi-object reinforcement learning. NeurIPS, 2023

  5. [13]

    Causal structure discovery from distributions arising from mixture dag

    Basil Saeed, Snigdha Panigrahi, and Caroline Uhler. Causal structure discovery from distributions arising from mixture dag. In ICML, 2020

  6. [14]

    An algorithm for fast recovery of sparse causal graphs

    Peter Spirtes and Clark Glymour. An algorithm for fast recovery of sparse causal graphs. Social science computer review , 9(1), 1991

  7. [15]

    Amortized inference for causal structure learning

    Lars Lorch, Scott Sussex, Jonas Rothfuss, Andreas Krause, and Bernhard Schölkopf. Amortized inference for causal structure learning. arXiv preprint arXiv:2205.12934, 2022

  8. [16]

    Dagma: Learning dags via m-matrices and a log-determinant acyclicity characterization

    Kevin Bello, Bryon Aragam, and Pradeep Ravikumar. Dagma: Learning dags via m-matrices and a log-determinant acyclicity characterization. Advances in Neural Information Processing Systems , 2022

  9. [17]

    Nonlinear learning method for local causal structures

    Xingyu Wu, Yan Zhong, Zhaolong Ling, Jie Yang, Li Li, Weiguo Sheng, and Bingbing Jiang. Nonlinear learning method for local causal structures. Information Sciences, 654:119789, 2024

  10. [18]

    Learning causal semantic representation for out-of-distribution prediction

    Chang Liu, Xinwei Sun, Jindong Wang, Haoyue Tang, Tao Li, Tao Qin, Wei Chen, and Tie-Yan Liu. Learning causal semantic representation for out-of-distribution prediction. Advances in Neural Information Processing Systems, 34:6155–6170, 2021

  11. [19]

    Invariant causal representation learning for out-of- distribution generalization

    Chaochao Lu, Yuhuai Wu, José Miguel Hernández-Lobato, and Bern- hard Schölkopf. Invariant causal representation learning for out-of- distribution generalization. In ICLR, 2021

  12. [20]

    Causpref: Causal preference learning for out-of-distribution recommendation

    Yue He, Zimu Wang, Peng Cui, Hao Zou, Yafeng Zhang, Qiang Cui, and Yong Jiang. Causpref: Causal preference learning for out-of-distribution recommendation. In Proceedings of the ACM Web Conference 2022 , pages 410–421, 2022

  13. [21]

    Counterfactual data augmentation using locally factored dynamics

    Silviu Pitis, Elliot Creager, and Animesh Garg. Counterfactual data augmentation using locally factored dynamics. Advances in Neural Information Processing Systems , 33:3976–3990, 2020

  14. [22]

    Seeing is not believing: Robust reinforcement learning against spurious correlation

    Wenhao Ding, Laixi Shi, Yuejie Chi, and Ding Zhao. Seeing is not believing: Robust reinforcement learning against spurious correlation. NeurIPS, 36:66328–66363, 2023

  15. [23]

    Mocoda: Model-based counterfactual data augmentation

    Silviu Pitis, Elliot Creager, Ajay Mandlekar, and Animesh Garg. Mocoda: Model-based counterfactual data augmentation. Advances in Neural Information Processing Systems , 35:18143–18156, 2022

  16. [24]

    Online dynamics learning for predictive control with an application to aerial robots

    Tom Z Jiahao, Kong Yao Chee, and M Ani Hsieh. Online dynamics learning for predictive control with an application to aerial robots. In Conference on Robot Learning , pages 2251–2261. PMLR, 2023

  17. [25]

    Dynamic learning from adaptive neural control of robot manipulators with prescribed performance

    Min Wang and Anle Yang. Dynamic learning from adaptive neural control of robot manipulators with prescribed performance. IEEE Transactions on Systems, Man, and Cybernetics: Systems , 2017

  18. [26]

    Dynamics learning with object-centric interaction networks for robot manipulation

    Jiayu Wang, Chuxiong Hu, Yunan Wang, and Yu Zhu. Dynamics learning with object-centric interaction networks for robot manipulation. IEEE Access, 2021

  19. [27]

    Real-time inverse dynamics learning for mus- culoskeletal robots based on echo state gaussian process regression

    Christoph Hartmann, Joschka Boedecker, Oliver Obst, Shuhei Ikemoto, and Minoru Asada. Real-time inverse dynamics learning for mus- culoskeletal robots based on echo state gaussian process regression. Robotics: Science and Systems VIII , pages 113–120, 2013

  20. [28]

    Cascaded gaus- sian processes for data-efficient robot dynamics learning

    Sahand Rezaei-Shoshtari, David Meger, and Inna Sharf. Cascaded gaus- sian processes for data-efficient robot dynamics learning. International Conference on Intelligent Robots and Systems , 2019

  21. [29]

    Introduction to dynamic systems: theory, models, and applications, volume 1

    David G Luenberger. Introduction to dynamic systems: theory, models, and applications, volume 1. Wiley New York, 1979

  22. [30]

    Parameter identification of robot dynamics

    Pradeep K Khosla and Takeo Kanade. Parameter identification of robot dynamics. In 1985 24th IEEE CDC , pages 1754–1760. IEEE, 1985

  23. [31]

    Differential equations, dynamical systems, and linear algebra

    Hirsch MH. Differential equations, dynamical systems, and linear algebra. Pure and Appl. Math. , 5, 1974

  24. [32]

    Bayesian structure learning with generative flow networks

    Tristan Deleu, António Góis, Chris Emezue, Mansi Rankawat, Simon Lacoste-Julien, Stefan Bauer, and Yoshua Bengio. Bayesian structure learning with generative flow networks. In Uncertainty in Artificial Intelligence, pages 518–528. PMLR, 2022

  25. [33]

    Bayesian learning of causal structure and mechanisms with gflownets and variational bayes

    Mizu Nishikawa-Toomey, Tristan Deleu, Jithendaraa Subramanian, Yoshua Bengio, and Laurent Charlin. Bayesian learning of causal structure and mechanisms with gflownets and variational bayes. arXiv preprint arXiv:2211.02763, 2022

  26. [34]

    Why did i fail? a causal- based method to find explanations for robot failures

    Maximilian Diehl and Karinne Ramirez-Amaro. Why did i fail? a causal- based method to find explanations for robot failures. IEEE Robotics and Automation Letters, 7(4):8925–8932, 2022

  27. [35]

    Do feature attribution methods correctly attribute features? In Proceedings of the AAAI Conference on Artificial Intelligence , 2022

    Yilun Zhou, Serena Booth, Marco Tulio Ribeiro, and Julie Shah. Do feature attribution methods correctly attribute features? In Proceedings of the AAAI Conference on Artificial Intelligence , 2022

  28. [36]

    The effectiveness of feature attribution methods and its correlation with automatic evaluation scores

    Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. Advances in Neural Information Processing Systems , 34:26422– 26436, 2021

  29. [37]

    Causal inference for de-biasing motion estimation from robotic observational data

    Junhong Xu, Kai Yin, Jason M Gregory, and Lantao Liu. Causal inference for de-biasing motion estimation from robotic observational data. arXiv preprint arXiv:2210.08679 , 2022

  30. [38]

    Review of causal discovery methods based on graphical models

    Clark Glymour, Kun Zhang, and Peter Spirtes. Review of causal discovery methods based on graphical models. Frontiers in genetics , 10:524, 2019

  31. [39]

    Causal imitation learning with unobserved confounders

    Junzhe Zhang, Daniel Kumor, and Elias Bareinboim. Causal imitation learning with unobserved confounders. NeurIPS, 2020

  32. [40]

    L. E. Dubins. On curves of minimal length with a constraint on average curvature, and with prescribed initial and terminal positions and tangents. American Journal of Mathematics , 79(3):497–516, 1957

  33. [41]

    Deep reinforcement learning in a handful of trials using probabilistic dynamics models

    Kurtland Chua, Roberto Calandra, Rowan McAllister, and Sergey Levine. Deep reinforcement learning in a handful of trials using probabilistic dynamics models. NeurIPS, 2018

  34. [42]

    An approximate epistemic uncertainty analysis approach in the presence of epistemic and aleatory uncertainties

    Eduard Hofer, Martina Kloos, Bernard Krzykacz-Hausmann, Jörg Peschke, and Martin Woltereck. An approximate epistemic uncertainty analysis approach in the presence of epistemic and aleatory uncertainties. Reliability Engineering & System Safety , 77(3):229–238, 2002

  35. [43]

    Task independent causal state abstraction

    Zizhao Wang, Xuesu Xiao, Yuke Zhu, and Peter Stone. Task independent causal state abstraction. In Robot Learning workshop , 2021

  36. [44]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017

  37. [45]

    Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor

    Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In ICML. PMLR, 2018

  38. [46]

    Addressing function approximation error in actor-critic methods

    Scott Fujimoto, Herke Hoof, and David Meger. Addressing function approximation error in actor-critic methods. In ICML. PMLR, 2018

  39. [47]

    Stable-baselines3: Reliable reinforcement learning implementations

    Antonin Raffin, Ashley Hill, Adam Gleave, Anssi Kanervisto, Max- imilian Ernestus, and Noah Dormann. Stable-baselines3: Reliable reinforcement learning implementations. JMLR, 2021

  40. [48]

    Lambert, and Roberto Calandra

    Luis Pineda, Brandon Amos, Amy Zhang, Nathan O. Lambert, and Roberto Calandra. Mbrl-lib: A modular library for model-based re- inforcement learning. Arxiv, 2021

  41. [49]

    Clearpath additional simulation worlds

    Clearpath Robotics. Clearpath additional simulation worlds. https:// github.com/clearpathrobotics/cpr_gazebo, 2020

  42. [50]

    Neural compression

    Matthew Muckley, Jordan Juravsky, Daniel Severo, Mannat Singh, Quentin Duval, and Karen Ullrich. Neural compression. https://github. com/facebookresearch/NeuralCompression

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.