Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Scalable Causal Imitation Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Causal imitation learning can be made tractable for long-horizon, high-dimensional control by replacing the full-horizon backdoor adjustment with a fixed-size sliding window, and then running off-policy soft Q-learning on the deconfounded r

desk verdict The empirical claim holds up, but Theorem 1's proof has a gap: Assumption 1(i) does not actually bound the temporal span of bidirected edges, so the windowed adjustment guarantee is unproven. read the letter →

arxiv 2607.17003 v1 pith:KRKJCTVV submitted 2026-07-18 cs.LG cs.AI

Eylam Tagor , Mingxuan Li , Elias Bareinboim This is my paper
classification cs.LGcs.AI
keywords causalimitationlearningunobservedconfoundingsequentialbackdoorcriterionsoftQ-learninglong-horizoncontrolslidingwindowadjustmentinversereinforcementdistributionshift
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 tries to establish that imitation learning from confounded demonstrations can scale to continuous-control tasks with horizons of a thousand or more steps. It argues that the sequential π-backdoor criterion, which is exact but intractable at scale, can be approximated by a fixed-size sliding window under an assumption of k-bounded, time-homogeneous confounding. With that windowed adjustment, two off-policy algorithms—Causal SQIL and Causal IQ-Learn—inherit the deconfounding guarantee while gaining the long-horizon stability of temporal-difference learning. A sympathetic reader would care because this is the first causal imitation framework that does not collapse on long-horizon, high-dimensional tasks: the paper reports that prior causal methods degrade sharply or reach zero success, while the proposed methods recover most of the expert's success rate.

What carries the argument

The load-bearing mechanism is the windowed sequential π-backdoor adjustment (Algorithm 1). It solves the exact sequential π-backdoor construction on a short proxy graph of horizon k+1, extracts per-action adjustment sets, then transfers them to the full horizon via time-homogeneity, clipping to a sliding window of width k. This converts an adjustment-set computation that grows linearly with the horizon into one of fixed size, and it determines which observed variables—and which lags—the imitator may condition on, excluding spurious colliders such as the wind-corrupted compass. The downstream SQIL and IQ-Learn objectives are unchanged; only the input representation is causally adjusted.

What would settle it

Construct a control environment with a persistent hidden goal or slowly accumulating external disturbance whose effect on actions grows beyond k steps; if a method using window k loses success while a method with a larger window or full history recovers it, the bounded-influence assumption is violated. Alternatively, build the graph from the paper's Example 3—where all edges span one step but the extended parent set reaches back two steps—and check that Algorithm 1's width-1 window leaves an open confounding path.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1: if the unrolled causal graph satisfies k-bounded time-homogeneous confounding—meaning every extended parent of any variable occurs within k timesteps and the local structure is time-invariant—then the windowed adjustment sets returned by Algorithm 1 satisfy the sequential π-backdoor criterion for the full-horizon graph. Consequently, feeding the encoded sliding-window representations into the standard SQIL and IQ-Learn objectives yields policies that are unbiased by unobserved confounders and learnable over long horizons. Empirically, on confounded maze-navigation tasks with 8- and 21-degree-of-freedom robots, Causal SQIL and Causal IQ-Learn reach roughly 90%

Load-bearing premise

The central premise is that all confounding influence on any action or state variable is confined to a fixed window of k timesteps and that this local causal structure is identical across time; if a confounder's effect lingers beyond the window, the adjusted sets may omit needed variables and confounding bias returns.

Editorial extensions

If this is right

  • Confounded imitation learning becomes feasible for horizons of 1000–2000 steps in continuous control, a regime where prior causal methods fail due to compounding error, sample inefficiency, or infeasible adjustment-set computation.
  • Causal adjustment is necessary for robustness: all non-causal baselines in the paper achieve near-zero success when the confounder distribution shifts at runtime, even when they look fine during training.
  • The causal adjustment layer is algorithm-agnostic: any future soft-Q-based imitation algorithm can be made confound-robust by feeding it the windowed causally-adjusted representation.
  • Under the bounded-influence assumption, conditioning on a fixed-size window rather than full history loses no causal information needed for deconfounding, and excessively large windows can hurt performance by bloating the representation.
  • On easier tasks the proposed methods sometimes surpass the expert, since off-policy Q-learning implicitly averages over disturbance distributions rather than committing to point estimates.

Reading between the lines

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

  • If the k-bounded influence assumption holds in a broader class of real systems with temporally localized disturbances, the same windowing trick could be applied to other causal structure-learning or deconfounding pipelines, reducing their computational cost from horizon-dependent to constant.
  • The paper's comparison suggests a testable division of labor: causal adjustment fixes confounding bias, while off-policy TD learning fixes compounding error; composing both is what enables scale, so one should expect either ingredient alone to fail on long-horizon confounded tasks.
  • For environments with persistent latent dependencies—hidden goals, slowly drifting dynamics—the fixed window will eventually be too small; a natural extension is to make the window size adaptive or to complement the variable selection with learned compression of long-range history, an idea the paper itself flags as future work.
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 / 5 minor

Summary. This paper introduces Causal SQIL and Causal IQ-Learn, two off-policy causal imitation learning algorithms that run soft Q-learning objectives on causally adjusted state representations. To make sequential π-backdoor adjustment tractable for long horizons, it proposes Algorithm 1, which computes adjustment sets on a short proxy graph of horizon k+1 and transfers them to the full graph as a fixed-size sliding window. The central theoretical claim is Theorem 1: under Assumption 1 (k-bounded time-homogeneous confounding), the transferred sets satisfy the sequential π-backdoor criterion. Empirically, the proposed algorithms are reported to retain substantial success on confounded AntMaze and HumanoidMaze tasks while non-causal methods collapse to near-zero success.

Significance. If the correctness of Theorem 1 can be established, the paper would be a valuable contribution: it combines causal adjustment with off-policy IL in a practical way and reduces the cost of sequential backdoor adjustment from O(H) to O(k). The empirical evaluation is thorough and well presented: it reports standard errors, ablates the window size, horizon, and confounding strength, compares against both causal and non-causal baselines, and includes an honest limitations section. The in-distribution invisibility of confounding demonstrated in Figure 4 is a particularly useful observation. However, the central correctness guarantee is currently false as stated, and this is a load-bearing issue: the theoretical basis for the windowed adjustment is not established by the proof supplied.

major comments (3)
  1. [Notation, Definition 3; Assumption 1(i); Appendix B, Step 1] Theorem 1 is false as stated. The formal definition pa+(A)=pa(C(A))\{A} does not imply that a variable B connected to A by a bidirected edge lies in pa+(A): being in the same C-component is not the same as being a parent of that C-component. Appendix B Step 1 asserts the converse ("for a bidirected edge B↔A with B earlier, B is collider-connected to A and hence B∈pa+"), which is invalid. Concretely, let k=1 and take variables B_{t-2}, C_{t-1}, S_t, S_{t+1}, action X_t, and terminal Y with edges B_{t-2}→C_{t-1}→S_t→S_{t+1}→Y, X_t→S_{t+1}→Y, and one bidirected edge X_t↔B_{t-2}. Then pa+(X_t)=∅, pa+(C_{t-1})={B_{t-2}} (lag 1), pa+(S_t)={C_{t-1}}, pa+(S_{t+1})={S_t,X_t}, pa+(Y)={S_{t+1}}, so Assumption 1(i) holds. The proxy graph of horizon k+1=2 omits B_{t-2}; Algorithm 1 returns Z_t=∅. But X_t is an ancestor of Y and the path X_t↔B_{t-2}→C_{t-1}→S_t→S_{t+1}→Y is open, so the sequential π-b
  2. [Appendix B, Step 1 ("sealing property")] The key step asserts that the window is "sealed against open escapes into the past" through a dichotomy of colliders/non-colliders and the Markov-boundary property. This is not derived formally from Assumption 1 or from Algorithm 1; it is the conclusion that must be proved. The counterexample in the previous comment satisfies all stated assumptions, yet the backdoor path leaves the window through a bidirected edge of span 2, so the sealing property fails. A rigorous proof needs either a graph-theoretic cut argument or a strictly stronger assumption that actually bounds C-component members and edge spans.
  3. [Assumption 1 and Section 3.2] Assumption 1(i) should be reformulated. To make Theorem 1 true, the bounded-influence condition must apply to all members of C(V_t) and their parents, or equivalently to the span of every directed and bidirected edge, not merely to pa(C(V_t))\{V_t}. The current definition is too weak, and the paper's own prose in the Notation section ("pa+ contains every variable collider-connected to V") is inconsistent with the formal definition. Additionally, the claim that the assumption is "naturally satisfied in MuJoCo" is not supported by any measurement; after strengthening the assumption, the paper should either provide evidence or present the windowed method as an approximation whose correctness is conditional on an explicit, verifiable stronger condition.
minor comments (5)
  1. [Section 2, Notation] The text says "Y∈U is the latent reward," but later Y is treated as an endogenous variable in V_H. Please clarify the formal status of Y (likely Y⊆V, with U reserved for exogenous noise).
  2. [Figure 5] The x-axis includes k=0, but Algorithm 1 requires k≥1 and Assumption 1 is stated for k≥1. Please explain how k=0 is defined or remove it from the ablation.
  3. [Algorithm 1 and Algorithm 3] The notation Pa+ and pa+ is used interchangeably, and the expression MB←Pa+(C(Ch+(OX)))\OX is difficult to parse. Please unify notation and define all operators when first used.
  4. [Section 3 and Table 1] The text reports normalized success rates (88.9%, 71.2%, 30.8%) for Causal BC while Table 1 reports raw success rates. Please make the distinction between raw and normalized values explicit in both places.
  5. [Appendix F.3] The runtime values appear to be for full-horizon FINDOX at small horizons, but the text says the runtime for H=2 and H=10 is "equivalent to that of Algorithm 1." This should be phrased more carefully: Algorithm 1 runs FINDOX on a proxy with horizon k+1, which is not literally the same as running FINDOX on a full graph with horizon H=k+1 if boundary conditions differ.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central guarantee is a conditional graph-transfer theorem, the algorithms compute rather than fit their inputs, and the self-citations are to independent prior graphical results.

full rationale

The derivation chain is not circular. Theorem 1 is a conditional statement: under Assumption 1 (k-bounded time-homogeneous confounding) and given that Algorithm 1's proxy-graph sets satisfy the sequential π-backdoor criterion in G^k, it proves those sets satisfy the criterion in the full graph G^H. The conclusion is not used to define Assumption 1 or the proxy admissibility premise; the proof is a transfer argument through a local subgraph isomorphism and an interface cut set. Algorithm 1 computes adjustment sets from a specified causal graph via FINDOX and the Markov boundary; no parameter is fitted to the reported success rates and then renamed a prediction. The window size k is a user-supplied hyperparameter, and Figure 5 reports sensitivity to it rather than treating it as a theoretically predicted value. The main external inputs (sequential π-backdoor criterion, FINDOX, Markov-boundary lemma) are cited to Kumor et al. (2021), whose authors overlap with the present paper, but this prior work is a peer-reviewed graphical result with stated assumptions that does not depend on the present paper's fitted values; it is therefore independent support rather than a circular self-reference. The empirical claims are measurements after training, not deductions from a fitted parameter. The paper's Appendix G explicitly limits the windowed approximation to bounded temporal influence and notes that long-range latent dependencies would violate Assumption 1; that is an honest limitation and a correctness condition, not circularity. The 'sealing property' in Appendix B, Step 1 may be under-justified under the formal definition of pa+(·), but an unsupported proof step is a correctness or assumption-strength concern, not a reduction of Theorem 1 to its own inputs. Thus no circular step is exhibited, and the appropriate score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central theorem rests on Assumption 1, a strong domain assumption about bounded temporal influence of confounders; the paper asserts it holds naturally for MuJoCo physics without measurement. Prior graphical results (sequential π-backdoor, FINDOX, projection) are imported as standard theory. No new theoretical entities are introduced; the latent wind/tremor and proxy sensors are benchmark constructs. The only free quantity materially affecting the claims is the window size k, selected per environment with documented sensitivity (Figure 5).

free parameters (1)
  • Sliding window size k = k=10 (AntMaze); k=2 (HumanoidMaze)
    k bounds the temporal reach of confounding in Assumption 1 and is chosen per environment. Figure 5 shows sensitivity: k=0 and k=100 degrade success rates substantially. No principled selection rule is provided; the correctness guarantee holds only for k that actually satisfies the assumption.
assumptions (5)
  • domain assumption Assumption 1 (k-Bounded Time-Homogeneous Confounding): every node in pa+(V_t) occurs no earlier than t−k; local causal structure among (V_{t−j},...,V_t, X_t) is isomorphic across t.
    Stated in Section 3.2 and used in Theorem 1. The paper justifies it for MuJoCo physics without measurement; if false, the windowed adjustment may omit confounders.
  • standard math Sequential π-backdoor criterion and FINDOX correctness (Kumor et al. 2021, Theorem 3.1, Definition 3.1, Lemma 3.2)
    Assumed as established background in Definition 2 and Algorithm 1 (reproduced as Algorithm 4).
  • standard math Latent projection onto V^O ∪ {Y} preserves d-separation and ancestral relationships (Verma & Pearl 1990; Tian & Pearl 2002)
    Used in Appendix B proof setup.
  • domain assumption SQIL and IQ-Learn objectives (soft Q-learning) converge to valid policies on the causally adjusted representation
    The paper applies off-the-shelf algorithms without new convergence theory; it assumes that the soft Bellman equations on (z,x) are well-posed.
  • domain assumption The causal graph G is known a priori and correctly specified
    Stated in Limitations (Appendix G). The method provides no procedure or guarantee for graph misspecification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable Causal Imitation Learning." pith.science (2026). https://pith.science/paper/KRKJCTVV

@misc{pith2026260717003,
  author       = {Pith},
  title        = {Pith review of: Scalable Causal Imitation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KRKJCTVV}},
  note         = {Machine review of arXiv:2607.17003}
}
abstract

Imitation learning enables learning a policy in an unknown environment with a latent reward signal using expert demonstrations, but it struggles when the imitator's and expert's observations are mismatched and unobserved confounders are present in expert demonstrations. By identifying appropriate adjustment sets via the sequential $\pi$-backdoor criterion, causal imitation learning (CIL) provides a framework for approximating the expert's policy from confounded data. However, existing CIL methods, Causal Behavioral Cloning (Causal BC) and Causal Generative Adversarial Imitation Learning (Causal GAIL), are designed for short-horizon, low-dimensional settings. When applied to continuous control tasks with long horizons and high-dimensional state-action spaces, these methods exhibit poor performance: Causal BC suffers from compounding errors, Causal GAIL is unstable and sample-inefficient, and sequential $\pi$-backdoor adjustment becomes impractical. We introduce Causal Soft Q Imitation Learning (SQIL) and Causal Inverse soft-Q Learning (IQ-Learn), two off-policy causal imitation learning algorithms that combine the causal adjustment framework with state-of-the-art inverse reinforcement learning objectives. Both algorithms operate on causally-adjusted state representations produced by an efficient approximation of the sequential $\pi$-backdoor criterion, exploiting the causal structure of continuous control environments to reduce the full-horizon adjustment to a fixed-size sliding window. We evaluate all methods in a suite of confounded environments and find that Causal SQIL and Causal IQ-Learn substantially outperform prior CIL algorithms on long-horizon tasks, sometimes surpassing the expert, whereas all causally unaware imitation methods fail to learn meaningful behavior.

Figures

Figures reproduced from arXiv: 2607.17003 by the authors.

Figure 1
Figure 1. Performance of various imitation learning algorithms on the Confounded AntMaze Large [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) Generic causal diagram for sequential imitation learning with unobserved confounders, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Causal BC on Confounded AntMaze. On Medium (left), the agent reaches the goal. On [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Evaluation return during training and runtime for GAIL, SQIL, and IQ-Learn (causal and [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Sensitivity of causal al￾gorithms to k in Algorithm 1 on Confounded AntMaze-Large. Windowed approximation is necessary [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Causal graph for Example 3, shown over the final three timesteps. Every edge spans a single timestep, yet the open path ρ, of XH ↔ XH−1 ← XH−2 → OH−1 ↔ OH ↔ Y , reaches the non-collider XH−2 at lag −2, outside the width-1 window WinH = {H − 1, H}. Example 3. Consider a…
Figure 7
Figure 7. Figure 7: Confounded AntMaze. P is global position, L is torso linear velocity, O is torso orienta￾tion, A is joint angles, T is torso angular velocity, J is joint angular velocities, U is the latent wind field, W is the compass, X is joint torques, and Y is the latent terminal …
Figure 8
Figure 8. Figure 8: Visualization for Confounded AntMaze, demonstrating the effect of latent winds [PITH_FULL_IMAGE:figures/full_fig_p029_8.png]
Figure 9
Figure 9. Figure 9: Confounded HumanoidMaze. P is global position, A is joint angles, H is head height, E is extremities in torso frame, V is torso vertical, C is center-of-mass gravity, J is joint velocities, U is the latent seismic tremor, W is the vibration sensor, X is joint torques, …
Figure 10
Figure 10. Figure 10: Visualization for Confounded HumanoidMaze, demonstrating the effect of latent seismic [PITH_FULL_IMAGE:figures/full_fig_p030_10.png]
Figure 11
Figure 11. Figure 11: Runtime to compute adjustment sets as a function of horizon [PITH_FULL_IMAGE:figures/full_fig_p034_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 2 canonical work pages

  1. [1]

    Causal Imitation Learning With Unobserved Confounders , url =

    Zhang, Junzhe and Kumor, Daniel and Bareinboim, Elias , booktitle =. Causal Imitation Learning With Unobserved Confounders , url =

  2. [2]

    Sequential Causal Imitation Learning with Unobserved Confounders , url =

    Kumor, Daniel and Zhang, Junzhe and Bareinboim, Elias , booktitle =. Sequential Causal Imitation Learning with Unobserved Confounders , url =

  3. [3]

    The Eleventh International Conference on Learning Representations , year=

    Causal Imitation Learning via Inverse Reinforcement Learning , author=. The Eleventh International Conference on Learning Representations , year=

  4. [4]

    Causal Imitation for Markov Decision Processes: a Partial Identification Approach , url =

    Ruan, Kangrui and Zhang, Junzhe and Di, Xuan and Bareinboim, Elias , booktitle =. Causal Imitation for Markov Decision Processes: a Partial Identification Approach , url =. doi:10.52202/079017-2781 , editor =

    work page doi:10.52202/079017-2781

  5. [5]

    Dragan and Sergey Levine , title =

    Siddharth Reddy and Anca D. Dragan and Sergey Levine , title =. 8th International Conference on Learning Representations,. 2020 , url =

    2020

  6. [6]

    IQ-Learn: Inverse soft-Q Learning for Imitation , url =

    Garg, Divyansh and Chakraborty, Shuvam and Cundy, Chris and Song, Jiaming and Ermon, Stefano , booktitle =. IQ-Learn: Inverse soft-Q Learning for Imitation , url =

  7. [7]

    2026 , eprint=

    Causal Flow Q-Learning for Robust Offline Reinforcement Learning , author=. 2026 , eprint=

    2026

  8. [8]

    The 25th International Conference on Autonomous Agents and Multi-Agent Systems , year=

    Confounding Robust Continuous Control via Automatic Reward Shaping , author=. The 25th International Conference on Autonomous Agents and Multi-Agent Systems , year=

Show all 60 references
  1. [9]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Confounding Robust Deep Reinforcement Learning: A Causal Approach , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  2. [10]

    2025 , url=

    Seohong Park and Kevin Frans and Benjamin Eysenbach and Sergey Levine , booktitle=. 2025 , url=

    2025

  3. [11]

    Causality: Models, Reasoning, and Inference , author =

  4. [12]

    and Ibeling, Duligur and Icard, Thomas , title =

    Bareinboim, Elias and Correa, Juan D. and Ibeling, Duligur and Icard, Thomas , title =. Probabilistic and Causal Inference: The Works of Judea Pearl , pages =. 2022 , isbn =

    2022

  5. [13]

    Causal Confusion in Imitation Learning , url =

    de Haan, Pim and Jayaraman, Dinesh and Levine, Sergey , booktitle =. Causal Confusion in Imitation Learning , url =

  6. [14]

    Imitation Is Not Enough: Robustifying Imitation with Reinforcement Learning for Challenging Driving Scenarios , year=

    Lu, Yiren and Fu, Justin and Tucker, George and Pan, Xinlei and Bronstein, Eli and Roelofs, Rebecca and Sapp, Benjamin and White, Brandyn and Faust, Aleksandra and Whiteson, Shimon and Anguelov, Dragomir and Levine, Sergey , booktitle=. Imitation Is Not Enough: Robustifying Im...

  7. [15]

    Proceedings of Robotics: Science and Systems (RSS) , year=

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author=. Proceedings of Robotics: Science and Systems (RSS) , year=

  8. [16]

    Robotics: Science and Systems , year =

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware , author =. Robotics: Science and Systems , year =

  9. [17]

    arXiv preprint arXiv:2307.15818 , year=

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control , author=. arXiv preprint arXiv:2307.15818 , year=

    arXiv

  10. [18]

    Figueiredo Prudencio, Rafael and Maximo, Marcos R. O. A. and Colombini, Esther Luna , year=. A Survey on Offline Reinforcement Learning: Taxonomy, Review, and Open Problems , volume=. IEEE Transactions on Neural Networks and Learning Systems , publisher=. doi:10.1109/tnnls.202...

    2023

  11. [19]

    Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics , pages =

    A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , author =. Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics , pages =. 2011 , url =

    2011

  12. [20]

    Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence , pages =

    Maximum Entropy Inverse Reinforcement Learning , author =. Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence , pages =. 2008 , url =

    2008

  13. [21]

    Generative Adversarial Imitation Learning , url =

    Ho, Jonathan and Ermon, Stefano , booktitle =. Generative Adversarial Imitation Learning , url =

  14. [22]

    International Conference on Learning Representations , year=

    Learning Robust Rewards with Adverserial Inverse Reinforcement Learning , author=. International Conference on Learning Representations , year=

  15. [23]

    arXiv preprint arXiv:2005.01643 , year =

    Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems , author =. arXiv preprint arXiv:2005.01643 , year =

    2005 arXiv

  16. [24]

    Justin Fu and Aviral Kumar and Ofir Nachum and George Tucker and Sergey Levine , year=. D4

  17. [25]

    2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

    MuJoCo: A physics engine for model-based control , author=. 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=. 2012 , organization=

    2012

  18. [26]

    GitHub repository , howpublished =

    Leurent, Edouard , title =. GitHub repository , howpublished =. 2018 , publisher =

    2018

  19. [27]

    International Conference on Machine Learning , pages =

    Addressing Function Approximation Error in Actor-Critic Methods , author =. International Conference on Machine Learning , pages =. 2018 , url =

    2018

  20. [28]

    Andrew and Abbeel, Pieter and Peters, Jan , year=

    Osa, Takayuki and Pajarinen, Joni and Neumann, Gerhard and Bagnell, J. Andrew and Abbeel, Pieter and Peters, Jan , year=. An Algorithmic Perspective on Imitation Learning , volume=. Foundations and Trends® in Robotics , publisher=. doi:10.1561/2300000053 , number=

    doi

  21. [29]

    and Khosravi, Abbas and Nahavandi, Saeid , journal=

    Zare, Maryam and Kebria, Parham M. and Khosravi, Abbas and Nahavandi, Saeid , journal=. A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges , year=

  22. [30]

    2024 , issue_date =

    Chen, Li and Wu, Penghao and Chitta, Kashyap and Jaeger, Bernhard and Geiger, Andreas and Li, Hongyang , title =. 2024 , issue_date =. doi:10.1109/TPAMI.2024.3435937 , journal =

    2024

  23. [31]

    Exploring the Limitations of Behavior Cloning for Autonomous Driving , year=

    Codevilla, Felipe and Santana, Eder and Lopez, Antonio and Gaidon, Adrien , booktitle=. Exploring the Limitations of Behavior Cloning for Autonomous Driving , year=

  24. [32]

    and Chen, Fang , journal=

    Zheng, Boyuan and Verma, Sunny and Zhou, Jianlong and Tsang, Ivor W. and Chen, Fang , journal=. Imitation Learning: Progress, Taxonomies and Challenges , year=

  25. [33]

    , booktitle =

    Pomerleau, Dean A. , booktitle =. ALVINN: An Autonomous Land Vehicle in a Neural Network , url =

  26. [34]

    , title =

    Abbeel, Pieter and Ng, Andrew Y. , title =. 2004 , isbn =. doi:10.1145/1015330.1015430 , booktitle =

    2004

  27. [35]

    International Conference on Machine Learning , pages =

    Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor , author =. International Conference on Machine Learning , pages =. 2018 , url =

    2018

  28. [36]

    arXiv preprint arXiv:2110.10819 , year=

    Shaking the foundations: delusions in sequence models for interaction and control , author=. arXiv preprint arXiv:2110.10819 , year=

    arXiv

  29. [37]

    Spencer and Sanjiban Choudhury and Arun Venkatraman and Brian D

    Jonathan C. Spencer and Sanjiban Choudhury and Arun Venkatraman and Brian D. Ziebart and J. Andrew Bagnell , title =. CoRR , volume =. 2021 , url =. 2102.02872 , timestamp =

    2021 arXiv

  30. [38]

    Fighting Copycat Agents in Behavioral Cloning from Observation Histories , url =

    Wen, Chuan and Lin, Jierui and Darrell, Trevor and Jayaraman, Dinesh and Gao, Yang , booktitle =. Fighting Copycat Agents in Behavioral Cloning from Observation Histories , url =

  31. [39]

    2023 , cdate=

    Samuel Pfrommer and Yatong Bai and Hyunin Lee and Somayeh Sojoudi , title=. 2023 , cdate=

    2023

  32. [40]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Causal Imitation Learning under Temporally Correlated Noise , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , volume =

    2022

  33. [41]

    and Wu, Steven Z

    Swamy, Gokul and Choudhury, Sanjiban and Bagnell, J. and Wu, Steven Z. , booktitle =. Sequence Model Imitation Learning with Unobserved Contexts , url =

  34. [42]

    Transactions on Machine Learning Research , issn=

    Deconfounding Imitation Learning with Variational Inference , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  35. [43]

    Workshop on Spurious Correlation and Shortcut Learning: Foundations and Solutions , year=

    A Unifying Framework for Causal Imitation Learning with Hidden Confounders , author=. Workshop on Spurious Correlation and Shortcut Learning: Foundations and Solutions , year=

  36. [44]

    CoRR , volume=

    Yan Zeng and Shenglan Nie and Feng Xie and Libo Huang and Peng Wu and Zhi Geng , title=. CoRR , volume=. 2025 , month=

    2025

  37. [45]

    Eighteenth National Conference on Artificial Intelligence , pages =

    Tian, Jin and Pearl, Judea , title =. Eighteenth National Conference on Artificial Intelligence , pages =. 2002 , isbn =

    2002

  38. [46]

    Scandinavian Journal of Statistics , volume =

    Richardson, Thomas , title =. Scandinavian Journal of Statistics , volume =

  39. [47]

    Proceedings of the Sixth Annual Conference on Uncertainty in Artificial Intelligence , pages =

    Verma, Thomas and Pearl, Judea , title =. Proceedings of the Sixth Annual Conference on Uncertainty in Artificial Intelligence , pages =. 1990 , isbn =

    1990

  40. [48]

    Proceedings of the National Academy of Sciences , volume =

    Elias Bareinboim and Judea Pearl , title =. Proceedings of the National Academy of Sciences , volume =. 2016 , doi =

    2016

  41. [49]

    Conference on Robot Learning , year=

    DART: Noise Injection for Robust Imitation Learning , author=. Conference on Robot Learning , year=

  42. [50]

    Advances in Neural Information Processing Systems , editor=

    Mitigating Covariate Shift in Imitation Learning via Offline Data With Partial Coverage , author=. Advances in Neural Information Processing Systems , editor=. 2021 , url=

    2021

  43. [51]

    Mitigating Covariate Shift in Behavioral Cloning via Robust Stationary Distribution Correction , url =

    Seo, Seokin and Lee, Byung-Jun and Lee, Jongmin and Hwang, HyeongJoo and Yang, Hongseok and Kim, Kee-Eung , booktitle =. Mitigating Covariate Shift in Behavioral Cloning via Robust Stationary Distribution Correction , url =. doi:10.52202/079017-3466 , editor =

    doi

  44. [52]

    Distributionally Robust Behavioral Cloning for Robust Imitation Learning , year=

    Panaganti, Kishan and Xu, Zaiyan and Kalathil, Dileep and Ghavamzadeh, Mohammad , booktitle=. Distributionally Robust Behavioral Cloning for Robust Imitation Learning , year=

  45. [53]

    Distributionally Robust Imitation Learning , url =

    Bashiri, Mohammad Ali and Ziebart, Brian and Zhang, Xinhua , booktitle =. Distributionally Robust Imitation Learning , url =

  46. [54]

    Robust Imitation Learning against Variations in Environment Dynamics

    Jongseong Chae and Seungyul Han and Whiyoung Jung and Myungsik Cho and Sungho Choi and Youngchul Sung. Robust Imitation Learning against Variations in Environment Dynamics. Proceedings of Machine Learning Research. 2022

    2022

  47. [55]

    Geon-Hyeong Kim and Seokin Seo and Jongmin Lee and Wonseok Jeon and HyeongJoo Hwang and Hongseok Yang and Kee-Eung Kim , booktitle=. Demo. 2022 , url=

    2022

  48. [56]

    International Conference on Learning Representations , year=

    Behavioral Cloning from Noisy Demonstrations , author=. International Conference on Learning Representations , year=

  49. [57]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Imitation Learning from Imperfect Demonstration , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , editor =

    2019

  50. [58]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Discriminator-Weighted Offline Imitation Learning from Suboptimal Demonstrations , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , editor =

    2022

  51. [59]

    The Tenth International Conference on Learning Representations,

    Guy Tennenholtz and Assaf Hallak and Gal Dalal and Shie Mannor and Gal Chechik and Uri Shalit , title =. The Tenth International Conference on Learning Representations,. 2022 , url =

    2022

  52. [60]

    Proceedings of the 38th International Conference on Machine Learning (ICML) , year =

    Jongmin Lee and Wonseok Jeon and Byung-Jun Lee and Joelle Pineau and Kee-Eung Kim , title =. Proceedings of the 38th International Conference on Machine Learning (ICML) , year =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.