Distributionally Robust and Safe Imitation Learning

Ahmed Aboudonia; Naira Hovakimyan

REVIEW 3 major objections 4 minor 20 references

Distributionally Robust and Safe Imitation Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Combining Taylor-series imitation learning with a distributionally robust Wasserstein-ball loss and a CVaR safety penalty yields policies robust to policy-induced and uncertainty-induced distribution shifts while honoring safety constraints

desk verdict The robust TaSIL loss is a plausible heuristic, but the CVaR safety bound in Eq. (14) scales the wrong way in δ_s, so the safety claim does not hold. read the letter →

arxiv 2607.13436 v1 pith:3OIEEXP6 submitted 2026-07-15 cs.LG cs.SYeess.SY

Ahmed Aboudonia , Naira Hovakimyan This is my paper

classification cs.LGcs.SYeess.SY

keywords imitationlearningdistributionallyrobustoptimizationTaylorseriesL1-DRACdistributionshiftconditionalvalue-at-risksafecontrolunmannedaerialvehicle

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 build an imitation learning framework that does not collapse when the learned policy deviates slightly from the expert (policy-induced shift) or when the environment itself changes (uncertainty-induced shift). It does this by wrapping the Taylor-series imitation learning objective in a robust optimization: instead of matching the expert only on expert trajectories, it matches the expert on worst-case distributions within a Wasserstein ball around those trajectories, and appends a CVaR-based penalty for one-step-ahead constraint violations. The authors argue that this combination inherits a robustness certificate from the L1-DRAC closed-loop bound and reduces to standard TaSIL when no uncertainty is present. Their UAV simulator shows the safe-robust policy tracking the expert in benign regions and deviating near unsafe ones, where plain TaSIL and TaSIL+L1-DRAC both fail. If correct, this is a path to train safety-aware imitators offline without needing expert demonstrations in dangerous states.

What carries the argument

The central object is the distributionally robust and safe TaSIL loss (15), which combines a Wasserstein-ball robust imitation term (10) with a CVaR safety penalty. The Wasserstein ball B_rho_L centered at the Dirac distribution of the expert trajectory contains all distributions within 2-Wasserstein distance rho_L; rho_L = rho + rho_a is obtained by composing an L1-DRAC closed-loop robustness bound rho with a discretization bound rho_a from Proposition 3.1. The safety term uses the identity (14), converting a chance constraint on the true system into a deterministic linear condition with margin ||c_j||_2 sqrt(delta_s/n_c) rho_L. The whole argument runs on the assumption that this Wasserstei

What would settle it

A concrete test: run the training procedure with rho_L set to a fraction of the validated value, or with an uncertainty disturbance whose realized 2-Wasserstein distance from the nominal trajectory exceeds rho_L; if the learned policy then violates the constraint or deviates beyond the certified gap, the central claim is falsified. Also, computing the empirical W2 between true and nominal trajectories across many Monte Carlo runs and checking whether it exceeds rho_L would settle the bound's validity.

Watch

Extended reading notes

Core claim

The central claim is that the loss (15) — the distributionally robust TaSIL term (10) plus a ReLU-penalized CVaR constraint term — produces a policy whose closed-loop distribution stays within a Wasserstein radius of the expert's nominal trajectory while the safety penalty blocks constraint violations. The robustness term is built by replacing the pointwise TaSIL loss with a worst-case expectation over distributions in a ball B_rho_L centered at the Dirac distribution of the expert state; the safety term uses the closed-form CVaR reformulation (14) derived from the same Wasserstein bound. The paper's simulation evidence shows a UAV with the robust-safe policy staying in the safe region under

Load-bearing premise

The load-bearing premise is that the Wasserstein distance from the true closed-loop distribution to the nominal expert-trajectory Dirac is no larger than the empirically chosen rho_L; if real shifts exceed that radius, the worst-case training set omits the actual danger and the safety penalty is too optimistic.

Editorial extensions

If this is right

If the loss works as described, imitators can be trained entirely offline on nominal expert demonstrations but deployed in environments with unmodeled dynamics or noise, because the training objective already covers worst-case distributions.
The reduction to standard TaSIL when uncertainty vanishes (Theorem 3.1) means the framework is a strict generalization; existing TaSIL guarantees and implementations carry over by setting the ball radius to zero.
The CVaR safety penalty gives a principled reason for a learned policy to deviate from an expert: imitation is encouraged only where the expert's behavior is feasible under uncertainty, and constraint violations are actively penalized.
The same Wasserstein bound rho_L is reused for both robustness and safety, so any improvement in the L1-DRAC bound directly tightens both the performance guarantee and the safety margin.

Reading between the lines

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

Editorial inference: the authors leave the co-design problem open — rho_L is chosen empirically, but a joint design where the L1-DRAC parameters and the policy are optimized together, e.g., via the co-design theory they cite, would close the loop between the robustness bound and the learned policy.
Editorial inference: the safety term uses one-step-ahead states from expert data points; a natural extension is to apply the CVaR penalty on the adversarial distributions sampled in the robust term, which the authors mention as an option but do not evaluate.
Editorial inference: if the Wasserstein-ball formulation is sound, it suggests a general recipe for safe imitation — wrap any well-behaved imitation loss in a distributionally robust minimization and add a chance-constraint penalty derived from the same uncertainty set; the UAV example is one instantiation.
Editorial inference: the Taylor expansion of the expert policy around the expert trajectory is an approximation that degrades far from the demonstrations; a testable improvement would be to use multiple expansion centers or a learned expert model.

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.

Desk Editor's Note

read the letter

First, what's actually new: the paper extends DRIP by reformulating the TaSIL objective as a Wasserstein-ball robust loss (10) and adding a CVaR-based safety penalty (15). The robust-loss idea is reasonable as a heuristic: instead of aligning with the expert only on the nominal trajectory, it aligns on worst-case distributions within a radius ρ_L. The Taylor approximation of the expert policy along perturbed states is openly a surrogate, and the paper is honest about that. Proposition 3.1 is simple and correct, and the triangle inequality getting ρ_L = ρ + ρ_a is fine. Credit also for acknowledging that ρ_L is chosen empirically and that the co-design with L1-DRAC is unresolved.

The load-bearing problem is equation (14). The stress-test note lands. The margin term ||c_j||√(δ_s/n_c) ρ_L goes to zero as δ_s → 0, which is backwards for a chance constraint: as the allowed risk shrinks, the bound should become more conservative, not less. The counterexample in the stress-test makes it concrete: a mixture distribution at Wasserstein distance exactly ρ_L satisfies the chance constraint at equality but has positive CVaR, while (14) evaluates to zero. So (14) is not a valid upper bound on the CVaR. This is not a minor gap; the 'safe' part of the framework has no certified meaning as written.

The experimental support is a single UAV simulation with one qualitative figure and no error bars or quantitative metrics. That would be thin even if the theory were correct. The paper's own admission that ρ_L is chosen empirically means the demonstration is a fit, not a falsifiable prediction.

I disagree with the reader's emphasis on circularity as the main weakness. Importing bounds from L1-DRAC and DRIP is standard practice when those references are solid, and the paper is upfront about the import. The empirical ρ_L is a genuine limitation but the paper doesn't hide it. The wrong δ_s scaling is the real flaw.

Who is this for? Someone working on distributionally robust imitation learning might find the robust-loss reformulation worth looking at, but the safety piece would need to be reworked. As is, the paper should not be accepted. It deserves a serious referee to document the flaw, but the right outcome is reject-and-resubmit: the core idea is worth a second attempt.

Referee Report

3 major / 4 minor

Summary. The paper proposes a distributionally robust and safe imitation learning framework that combines Taylor Series Imitation Learning (TaSIL) with L1-Distributionally Robust Adaptive Control (L1-DRAC). The robust loss (10) penalizes policy discrepancy under a Wasserstein ambiguity set centered at the expert trajectory, and a CVaR safety penalty (15) is added to account for chance constraints. The method is evaluated on a UAV simulation in which the proposed policy is claimed to avoid unsafe regions under uncertainty.

Significance. The motivating problem is important: imitation learners operating under uncertainty can violate safety constraints. The architecture is plausible and builds on a strong line of prior work, and Proposition 3.1 is a reasonable elementary Lipschitz-propagation bound, modulo a square-root issue noted below. The authors are transparent about the empirical choice of rho_L and about the co-design limitation. However, the central safety bound (14) is invalid, and the robust loss (10) is not connected to a certified bound on the true distribution. As a result, the main claims of 'distributionally robust' and 'safe' are not substantiated. The simulation is qualitative and does not provide a falsifiable test of the safety mechanism.

major comments (3)

[§III.C, Eq. (14)] The closed-form CVaR bound has the wrong dependence on delta_s. The margin term ||c_j||_2 sqrt(delta_s/n_c) rho_L shrinks as delta_s -> 0, whereas a worst-case CVaR bound over a Wasserstein ball must grow as rho_L/sqrt(delta_s). Concretely, take scalar c=1, d=0, z_k=-0.1, rho_L=1, delta_s=0.01, and Q=(1-delta_s) delta_{-0.1}+delta_s delta_{9.9}. Then W2(Q, delta_{-0.1})=1=rho_L, so Q is in the ambiguity set, P(Q>0)=delta_s is exactly at the chance-constraint limit, but CVaR_{1-delta_s}(Q)=9.9>0. Eq. (14) evaluates to -0.1+0.1=0<=0, declaring the distribution safe even though the CVaR constraint (11) is violated. Thus (14) is not a valid upper bound, and the safety loss (15) can certify unsafe policies. The risk-allocation factor delta_s/n_c should appear in the denominator (or be justified by a proper union-bound/Lipschitz argument).
[§III.B, Prop. 3.1 and Eq. (10)] The distributionally robust objective is not connected to a valid bound. Proposition 3.1 bounds W2(X*_k(xi, pi_hat^s_p), delta_{x_k(xi, pi_hat^s_p)}), but Eq. (10) centers the ambiguity set at the Dirac of the expert trajectory x_k(xi, pi*). The text acknowledges this and says rho_L can be enlarged to account for the approximation error, but no bound on that enlargement is provided. Similarly, the discretized systems are assumed to inherit the continuous-time Wasserstein bound with the same rho after acknowledging that this may not hold. Thus Eq. (10) is a heuristic surrogate, not a certified distributionally robust loss. In addition, Proposition 3.1 states a bound on W2, but its proof sets rho_a = max_k E[||e_k||^2] while W2^2 = E[||e_k||^2]; the theorem statement and proof disagree by a square root, and this carries into the definition rho_L = rho + rho_a.
[§IV] The empirical demonstration is qualitative and does not validate the safety claim. The only success metric is a visual inspection of Fig. 1; there are no Monte Carlo statistics for constraint violations, imitation gaps, or risk-level sensitivity. Since rho_L is chosen empirically (end of §III.C), the favorable bottom-right panel is consistent with tuning rather than a falsifiable prediction. The paper's statement that the case study 'demonstrates the effectiveness' of the framework is therefore overstated, especially because the safety mechanism itself is based on the invalid bound (14).

minor comments (4)

[§III.B] Typo: 'Euler-Marumaya' should be 'Euler-Maruyama'.
[§III.A and §IV] The symbol lambda is used both for the L1-DRAC adaptation parameter (lambda_s in §III.A) and for the safety weight in Eq. (13). This is confusing; please use distinct symbols.
[§IV] The safe region is defined as {(x,y) in R^2 | x <= 0.8, y <= 0.8}, which is unbounded and does not match the narrative of 'avoiding unsafe regions' along the top-right portion of the circular trajectory. Please clarify the intended unsafe set.
[Theorem 3.1] Theorem 3.1 is essentially a tautology: without uncertainty, the ambiguity set collapses to the Dirac measure. If kept, it should be framed as a sanity check rather than a robustness result.

Circularity Check

2 steps flagged · score 6.0 of 10

Central robust/safe claim rests on a self-cited Wasserstein bound and on an empirically chosen radius ρ_L, so the Fig. 1 demonstration is a fitted demonstration rather than an independent prediction.

self citation load bearing [Section III.A (Robustness), after Eq. (6)]
"Notice that, given a policy π and its corresponding L1-DRAC-augmented policy π^l, it is shown in [17] that there exists a scalar ρ≥0 such that W2(Xt(ξ,π^l), X*_t(ξ,π))≤ρ for all t∈[0,Th], under mild regularity assumptions on the nominal dynamics and the uncertain terms in the true system... The specific case in which π is the standard TaSIL policy is addressed in [9]."

The distributionally robust loss (10) and the safety loss (14)-(15) are built on the Wasserstein ball B_{ρ_L}, whose radius ρ_L = ρ + ρ_a depends critically on the existence of this bound. The bound is not derived in the present paper; it is imported from [17] and [9], and both references share authors with the present paper (Hovakimyan in both, Aboudonia in [9]). The paper does not reproduce the bound, state and verify its assumptions, or test it against an external benchmark, so the central premise of the claimed robustness is a load-bearing self-citation.
fitted input called prediction [Section III.C (Safety), final paragraph; used in Section IV simulations]
"The proposed approach employs the bound ρ_L = ρ + ρ_a for learning the TaSIL policy, as in (15). The computation of the L1-DRAC bound, however, depend on the nominal dynamics under the learned policy, leading to a coupled co-design problem. ... In the present work, ρ_L is chosen empirically."

The safe/robust policy shown in Fig. 1 is trained using (15) with a hand-chosen ρ_L. Since ρ_L directly sets the size of the Wasserstein ball in (10) and the safety margin in (14), the observed safety and robustness behavior is a direct consequence of the chosen parameter, not a parameter-free prediction of the framework. The paper presents the empirical ρ_L as a substitute for the unavailable bound, so the simulation reduces to a calibration demonstration rather than an independent validation of the claimed robustness certificate.

full rationale

The proposed loss formulation itself is not circular: it combines TaSIL, L1-DRAC, and CVaR penalties in a well-defined objective, and Theorem 3.1 is only a consistency check showing that the new loss reduces to standard TaSIL without uncertainties. However, the paper's central claim—that the framework is distributionally robust and safe—depends on (i) a Wasserstein bound whose existence is imported from prior self-cited work and (ii) an empirically chosen ρ_L. The Fig. 1 demonstration therefore shows the effect of a fitted parameter, not a falsifiable independent prediction. I do not count the apparent δ_s-scaling problem in Eq. (14) as circularity; that is a correctness risk. Given the load-bearing self-citation and the empirically fitted robustness radius, a score of 6 is appropriate for partial circularity.

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

The central robustness/safety claims rest on the empirically chosen ρ_L and imported bounds from same-group papers. None of the theoretical bounds are computed; the safety formula includes an unverified—and likely incorrect—dependence on the risk level δ_s.

free parameters (5)

ρ_L = not reported; chosen empirically
Controls the Wasserstein ball in the robust loss (10) and the safety margin in (14). Paper states it is chosen empirically.
Safety penalty weight λ = not reported
Weight of the ReLU safety penalty in (13)/(15); not specified in simulation section.
L1-DRAC parameters ω, T_s, λ_s = ω=5, T_s=0.01, λ=5
Adaptive controller tuning; may affect the W2 bound ρ.
Risk parameter δ_s = not reported
Allowed chance-constraint violation probability; not given in the simulation.
Lipschitz constant L of f_cl = not estimated
Assumed finite in Proposition 3.1; not computed for the learned policy.

assumptions (5)

domain assumption L1-DRAC bound W2(X_t(ξ,π^l), X*_t(ξ,π)) ≤ ρ from [17]
Imported from same-group prior work; regularity assumptions not restated.
ad hoc to paper Discretized systems inherit the same Wasserstein bound with the same ρ
Paper says 'we maintain the constant ρ and assume' the bound holds for discrete-time.
domain assumption f_cl(x)=Ax+Bπ^s_p(x) is Lipschitz with constant L
Used in Prop 3.1; not verified for the trained neural network.
ad hoc to paper Closed-form worst-case CVaR bound (14) from [18] is applicable with the stated δ_s dependence
No derivation given; the term ||c||√(δ_s/n_c)ρ_L has the wrong monotonicity in δ_s for a chance constraint.
domain assumption Expert policy and its derivatives are available along expert trajectories
Standard TaSIL assumption, used in Taylor expansion.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distributionally Robust and Safe Imitation Learning." pith.science (2026). https://pith.science/paper/3OIEEXP6

@misc{pith2026260713436,
  author       = {Pith},
  title        = {Pith review of: Distributionally Robust and Safe Imitation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3OIEEXP6}},
  note         = {Machine review of arXiv:2607.13436}
}

read the original abstract

Imitation learning (IL) has achieved remarkable success in complex decision-making tasks. However, its performance is highly sensitive to distribution shifts, which can pose significant safety risks. We propose a distributionally robust and safe IL framework that explicitly addresses both policy-induced and uncertainty-induced distribution shifts. Our approach develops a unified framework leveraging Taylor Series Imitation Learning (TaSIL) to mitigate policy-induced shifts and distributionally robust adaptive control to handle uncertainty-induced shifts. This architecture enables the formulation of an IL problem that optimizes performance under distributional uncertainty while systematically accounting for safety constraints. We demonstrate the effectiveness of the proposed approach on an unmanned aerial vehicle (UAV) case study where the UAV performs a task in an uncertain environment while avoiding unsafe regions.

Figures

Figures reproduced from arXiv: 2607.13436 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 4 linked inside Pith

[1]

Recent advances in robot learning from demonstration,

H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,”Annual review of control, robotics, and autonomous systems, vol. 3, no. 1, pp. 297– 330, 2020

2020
[2]

A survey of imitation learning: Algorithms, recent developments, and challenges,

M. Zare, P. M. Kebria, A. Khosravi, and S. Nahavandi, “A survey of imitation learning: Algorithms, recent developments, and challenges,” IEEE Transactions on Cybernetics, vol. 54, no. 12, pp. 7173–7186, 2024

2024
[3]

Survey of imitation learning for robotic manipulation: B. fang et al.,

B. Fang, S. Jia, D. Guo, M. Xu, S. Wen, and F. Sun, “Survey of imitation learning for robotic manipulation: B. fang et al.,”Interna- tional Journal of Intelligent Robotics and Applications, vol. 3, no. 4, pp. 362–369, 2019

2019
[4]

A survey on imitation learning techniques for end-to-end autonomous vehicles,

L. Le Mero, D. Yi, M. Dianati, and A. Mouzakitis, “A survey on imitation learning techniques for end-to-end autonomous vehicles,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 9, pp. 14128–14147, 2022

2022
[5]

Imitation learning for legged robot loco- motion: a survey,

K. Z. Mirza and S. Singh, “Imitation learning for legged robot loco- motion: a survey,”Frontiers in Robotics and AI, vol. 12, p. 1678567, 2025

2025
[6]

Recent advances in imitation learning from observation,

F. Torabi, G. Warnell, and P. Stone, “Recent advances in imitation learning from observation,”arXiv preprint arXiv:1905.13566, 2019

arXiv 1905
[7]

Data quality in imitation learning,

S. Belkhale, Y . Cui, and D. Sadigh, “Data quality in imitation learning,”Advances in neural information processing systems, vol. 36, pp. 80375–80395, 2023

2023
[8]

Efficient deep learning of robust policies from mpc using imitation and tube-guided data augmentation,

A. Tagliabue and J. P. How, “Efficient deep learning of robust policies from mpc using imitation and tube-guided data augmentation,”IEEE Transactions on Robotics, vol. 40, pp. 4301–4321, 2024

2024

Show all 20 references

[9]

Distributionally robust imi- tation learning: Layered control architecture for certifiable autonomy,

A. Gahlawat, A. Aboudonia, S. Banik, N. Hovakimyan, N. Matni, A. Ames, G. Zardini, and A. Speranzon, “Distributionally robust imi- tation learning: Layered control architecture for certifiable autonomy,” inAIAA SCITECH 2026 Forum, p. 2169, 2026

2026
[10]

A reduction of imitation learning and structured prediction to no-regret online learning,

S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” inProceedings of the fourteenth international conference on artificial intelligence and statistics, pp. 627–635, JMLR Workshop and Conference Proceed...

2011
[11]

Dart: Noise injection for robust imitation learning,

M. Laskey, J. Lee, R. Fox, A. Dragan, and K. Goldberg, “Dart: Noise injection for robust imitation learning,” inConference on robot learning, pp. 143–156, PMLR, 2017

2017
[12]

A survey of inverse reinforcement learning: Challenges, methods and progress,

S. Arora and P. Doshi, “A survey of inverse reinforcement learning: Challenges, methods and progress,”Artificial Intelligence, vol. 297, p. 103500, 2021

2021
[13]

Generative adversarial imitation learning,

J. Ho and S. Ermon, “Generative adversarial imitation learning,” Advances in neural information processing systems, vol. 29, 2016

2016
[14]

Tasil: Taylor series im- itation learning,

D. Pfrommer, T. Zhang, S. Tu, and N. Matni, “Tasil: Taylor series im- itation learning,”Advances in Neural Information Processing Systems, vol. 35, pp. 20162–20174, 2022

2022
[15]

Robust imitation learning against variations in environment dynamics,

J. Chae, S. Han, W. Jung, M. Cho, S. Choi, and Y . Sung, “Robust imitation learning against variations in environment dynamics,” inIn- ternational Conference on Machine Learning, pp. 2828–2852, PMLR, 2022

2022
[16]

Distribution- ally robust behavioral cloning for robust imitation learning,

K. Panaganti, Z. Xu, D. Kalathil, and M. Ghavamzadeh, “Distribution- ally robust behavioral cloning for robust imitation learning,” in2023 62nd IEEE Conference on Decision and Control (CDC), pp. 1342– 1347, IEEE, 2023

2023
[17]

L 1- DRAC: Distributionally robust adaptive control,

A. Gahlawat, S. H. Karumanchi, and N. Hovakimyan, “L 1- DRAC: Distributionally robust adaptive control,”arXiv preprint arXiv:2509.04619, 2025

2025 arXiv
[18]

Wasserstein distributionally robust adap- tive covariance steering,

A. Gahlawat, V . Khatana, D. Wang, S. H. Karumanchi, N. Hov- akimyan, and P. V oulgaris, “Wasserstein distributionally robust adap- tive covariance steering,” in2025 IEEE 64th Conference on Decision and Control (CDC), pp. 4845–4850, IEEE, 2025

2025
[19]

Certifying some distributional robustness with principled adversarial training,

A. Sinha, H. Namkoong, R. V olpi, and J. Duchi, “Certifying some distributional robustness with principled adversarial training,”arXiv preprint arXiv:1710.10571, 2017

2017 arXiv
[20]

A mathematical theory of co-design,

A. Censi, “A mathematical theory of co-design,”arXiv preprint arXiv:1512.08055, 2015

2015 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Recent advances in robot learning from demonstration,

H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,”Annual review of control, robotics, and autonomous systems, vol. 3, no. 1, pp. 297– 330, 2020

2020

[2] [2]

A survey of imitation learning: Algorithms, recent developments, and challenges,

M. Zare, P. M. Kebria, A. Khosravi, and S. Nahavandi, “A survey of imitation learning: Algorithms, recent developments, and challenges,” IEEE Transactions on Cybernetics, vol. 54, no. 12, pp. 7173–7186, 2024

2024

[3] [3]

Survey of imitation learning for robotic manipulation: B. fang et al.,

B. Fang, S. Jia, D. Guo, M. Xu, S. Wen, and F. Sun, “Survey of imitation learning for robotic manipulation: B. fang et al.,”Interna- tional Journal of Intelligent Robotics and Applications, vol. 3, no. 4, pp. 362–369, 2019

2019

[4] [4]

A survey on imitation learning techniques for end-to-end autonomous vehicles,

L. Le Mero, D. Yi, M. Dianati, and A. Mouzakitis, “A survey on imitation learning techniques for end-to-end autonomous vehicles,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 9, pp. 14128–14147, 2022

2022

[5] [5]

Imitation learning for legged robot loco- motion: a survey,

K. Z. Mirza and S. Singh, “Imitation learning for legged robot loco- motion: a survey,”Frontiers in Robotics and AI, vol. 12, p. 1678567, 2025

2025

[6] [6]

Recent advances in imitation learning from observation,

F. Torabi, G. Warnell, and P. Stone, “Recent advances in imitation learning from observation,”arXiv preprint arXiv:1905.13566, 2019

arXiv 1905

[7] [7]

Data quality in imitation learning,

S. Belkhale, Y . Cui, and D. Sadigh, “Data quality in imitation learning,”Advances in neural information processing systems, vol. 36, pp. 80375–80395, 2023

2023

[8] [8]

Efficient deep learning of robust policies from mpc using imitation and tube-guided data augmentation,

A. Tagliabue and J. P. How, “Efficient deep learning of robust policies from mpc using imitation and tube-guided data augmentation,”IEEE Transactions on Robotics, vol. 40, pp. 4301–4321, 2024

2024