Action-Constrained Imitation Learning (ACIL)

• ACIL is an imitation-learning paradigm where the learner, constrained by a limited action set, mimics an expert whose actions exceed its feasibility.
• It addresses occupancy measure distortion by shifting focus from direct action replication to matching state trajectories using methods like DTWIL with MPC and CEM.
• Extensions such as constraint-aware behavior cloning and SCOPIL incorporate safety penalties and dual descent objectives to ensure feasible, robust policy performance.

to=arxiv_search.^^^^^^^^7query7^^^^^^^^ ￣奇米json {"^^^^^^^^7query7 Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^^^^^", "^^^^^^^^7max_results7 ^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7query7^^^^} to=arxiv_search.^^^^^^^^7query7^^^^^^^^ ോഗjson {"^^^^^^^^7query7 Imitation Learning\" OR ^^^^^^^^7ti:\7 Simple Approach to Constraint-Aware Imitation Learning with Application to Autonomous Racing\" OR ^^^^^^^^7ti:\7 safe, constrained policies via imitation learning: Connection to Probabilistic Inference and a Naive Algorithm\"", "^^^^^^^^7max_results7 ^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7query7^^^^} Action-Constrained Imitation Learning (ACIL) denotes an imitation-learning setting in which the imitator must learn from demonstrations generated by an expert with a larger effective action capability, while the imitator itself is restricted to a feasible action set. In the formalization introduced in ^^^^7max_results7query7max_results7 OR ti:\7^^^^, the central difficulty is not merely that some expert actions are infeasible, but that infeasible actions induce a mismatch in future state visitation and therefore a mismatch in occupancy measure. ACIL accordingly shifts the learning target away from exact state-action reproduction and toward state trajectories that remain as similar as possible to the expert’s under the imitator’s action constraints (&&&^^^^7query7^^^^&&&).

^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^. Formal definition and problem structure

The strict ACIL formulation models the environment as an action-constrained MDP,

PRESERVED_PLACEHOLDER_^^^^7query7^^^^

where PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ is the feasible action set at state PRESERVED_PLACEHOLDER_^^^^7max_results7^^^^, and the imitator must satisfy PRESERVED_PLACEHOLDER_^^^^7query7^^^^. The expert policy PRESERVED_PLACEHOLDER_^^^^7ti:\7^^^^ acts in the ambient action space PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7^^^^ and may select actions outside PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7^^^^ for some states. The resulting objective is not exact state-action imitation but minimization of discrepancy between the learner’s and the expert’s state sequences:

$$\pi^*:=\underset{\pi\in \Pi_{\mathcal{C}}}{\arg\min}\; \mathbb{E}_{\tau_e\sim \pi_e,\;\tau\sim \pi} \big[d(\sigma(\tau), \sigma(\tau_e)) \mid s_0\sim \mu\big].$$

Here $\sigma(\tau)$ denotes the state sequence of trajectory $\tau$, and PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7query7^^^^ is allowed to compare sequences of different lengths (&&&^^^^7query7^^^^&&&).

This formulation makes the action-capability gap explicit. The expert and imitator share the same state and action semantics, but the imitator is the weaker agent because it must remain in PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^. A constrained learner may therefore fail to reproduce not only expert actions but also the expert’s timing and downstream state distribution. In this sense, ACIL differs from ordinary learning from demonstrations, where realizability of the demonstrator’s behavior is usually assumed.

Related ^^^^7max_results7query7max_results7 OR ti:\7^^^^ work broadens the practical scope of the topic without adopting exactly the same formal definition. One line treats constraint-aware imitation as behavioral cloning augmented with a learned safety penalty based on predicted next-state feasibility (&&&^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^&&&). Another line treats constraints as implicit in demonstrations from a constraint-abiding expert and learns a maximum-entropy policy with a KL-based imitation term and dual gradient descent (&&&^^^^7max_results7^^^^&&&). This suggests a useful distinction between ACIL in the narrow sense of explicit feasible action sets and ACIL in a broader sense of imitation under feasibility or safety restrictions.

^^^^7max_results7^^^^. Occupancy measure distortion and why standard imitation learning fails

The defining conceptual claim of ACIL is that action constraints create an unavoidable occupancy-measure mismatch. If

PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7max_results7^^^^

then any feasible learner must satisfy PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7query7^^^^ for PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7ti:\7^^^^, while the expert may have PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7 OR ti:\7^^^^ on those same state-action pairs. Exact occupancy matching is therefore infeasible in general. The ^^^^7max_results7query7max_results7 OR ti:\7^^^^ ACIL paper terms the resulting phenomenon occupancy measure distortion (&&&^^^^7query7^^^^&&&).

A common constrained-RL baseline is to compute an unconstrained action PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7 OR ti:\7^^^^ and project it to feasibility with PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)77^^^^. ACIL argues that such projection does not resolve the imitation problem, because projected actions alter the next state, and the induced state deviation compounds over time. The mismatch is therefore trajectory-level rather than timestep-local. The Maze^^^^7max_results7^^^^D example described in the paper makes this concrete: a learner that takes the “closest” feasible turn still turns too late, collides, and departs from the intended route, whereas a temporally warped but feasible path can remain behaviorally similar (&&&^^^^7query7^^^^&&&).

This diagnosis explains why several standard imitation-learning families become ill-posed under action constraints. Behavior cloning regresses toward infeasible expert actions and suffers amplified covariate shift once executed actions are projected or clipped. Adversarial imitation methods such as GAIL attempt to match an unreachable expert occupancy measure. Inverse RL can recover a reward, but if the learner’s feasible set differs from the expert’s, the same reward need not produce the same occupancy measure. State-only methods avoid direct action mismatch, yet they still try to imitate trajectories generated by unconstrained actions and therefore do not eliminate the capability-gap problem. The common failure mode is not merely lack of safety, but mismatch between the target behavior class and the learner’s feasible behavior class.

^^^^7query7^^^^. DTWIL and surrogate action-constrained demonstrations

The principal algorithm proposed for strict ACIL is DTWIL (Dynamic Time Warping Imitation Learning). Its strategy is to replace infeasible expert demonstrations with surrogate, action-constrained demonstrations before running a standard imitation learner. The method decomposes the problem into two stages: first, generate a surrogate dataset PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^8 consisting of feasible trajectories that start from the same initial states and remain close to the expert’s state trajectories; second, train any off-the-shelf IL method on those feasible surrogates (&&&^^^^7query7^^^^&&&).

For an expert trajectory PRESERVED_PLACEHOLDER_^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^9, surrogate construction is posed as

PRESERVED_PLACEHOLDER_^^^^7max_results7query7^^^^

where PRESERVED_PLACEHOLDER_^^^^7max_results7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ is the set of feasible trajectories. The discrepancy is instantiated with Dynamic Time Warping (DTW) rather than pointwise PRESERVED_PLACEHOLDER_^^^^7max_results7max_results7^^^^ matching. DTW is critical because constrained learners often require more steps than the expert to realize a comparable path. The DTW distance is defined through a monotone alignment path over the pairwise distance matrix and is computed by the standard dynamic-programming recursion (&&&^^^^7query7^^^^&&&).

Direct optimization over variable-length feasible trajectories is handled through Model Predictive Control (MPC). At each surrogate timestep, MPC plans over horizon PRESERVED_PLACEHOLDER_^^^^7max_results7query7^^^^, rolls out candidate feasible action sequences under a learned forward model, and scores them by DTW against the current segment of the expert trajectory. Only the first action is executed, after which planning is repeated. Candidate sequences are optimized with the Cross-Entropy Method (CEM), and feasibility is enforced during sampling through rejection sampling. A learned ensemble of probabilistic neural dynamics models, with ensemble size PRESERVED_PLACEHOLDER_^^^^7max_results7ti:\7^^^^, supplies the predictive model used during planning (&&&^^^^7query7^^^^&&&).

Two implementation mechanisms are especially central. The first is the progression parameter PRESERVED_PLACEHOLDER_^^^^7max_results7 OR ti:\7^^^^, which tracks how far the surrogate has progressed along the expert trajectory. It advances asynchronously according to the DTW warping path rather than synchronizing one learner step to one expert step. The second is Expert Regularized Control (ERC), used in delicate environments such as Hopper, where early planned actions are blended with projected expert actions. The paper reports that replanning every step is better than executing several open-loop MPC steps, and that asynchronous progression is materially better than synchronous progression on Hopper (&&&^^^^7query7^^^^&&&).

The downstream significance of DTWIL is architectural rather than merely procedural. It decouples constraint satisfaction from policy fitting. Once the surrogate trajectories are feasible, the imitation learner no longer targets an impossible expert occupancy measure but a realizable surrogate occupancy that preserves the expert’s state-space behavior up to temporal warping.

^^^^7ti:\7^^^^. Constraint-aware and safety-aware extensions

Two closely related ^^^^7max_results7query7max_results7 OR ti:\7^^^^ methods address similar concerns from different angles. Constraint-aware Behavior Cloning augments behavioral cloning with a learned safety penalty derived from a differentiable approximation of a one-step predictive safety filter. The setup considers a nonlinear deterministic discrete-time system with partial observation,

PRESERVED_PLACEHOLDER_^^^^7max_results7 OR ti:\7^^^^

subject to PRESERVED_PLACEHOLDER_^^^^7max_results77^^^^ and PRESERVED_PLACEHOLDER_^^^^7max_results7^^^^8. The ideal imitation objective explicitly requires

PRESERVED_PLACEHOLDER_^^^^7max_results7^^^^9

where PRESERVED_PLACEHOLDER_^^^^7query7query7^^^^ denotes the backward reachable tube of states from which admissible control can reach the target set while satisfying constraints. Because the dynamics and safe set are assumed unavailable, the method learns a surrogate dynamics model PRESERVED_PLACEHOLDER_^^^^7query7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ and a safety classifier PRESERVED_PLACEHOLDER_^^^^7query7max_results7^^^^, leading to the final objective

PRESERVED_PLACEHOLDER_^^^^7query7query7^^^^

The method uses DAgger-style data aggregation, a safety auto-labeling heuristic based on successful and failed rollouts, local convex-hull filtering for ambiguous failed states, and updates the dynamics model every PRESERVED_PLACEHOLDER_^^^^7query7ti:\7^^^^ iterations and the safety classifier every PRESERVED_PLACEHOLDER_^^^^7query7 OR ti:\7^^^^ iterations. Crucially, “During test time, only the policy network PRESERVED_PLACEHOLDER_^^^^7query7 OR ti:\7^^^^ is used, while all other modules participate solely in training” (&&&^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^&&&).

SCOPIL addresses constrained behavior through a different formal route. It assumes a discounted MDP and a dataset of demonstrations from a constraint-abiding expert, but does not assume explicit access to the constraint costs. Instead, it formulates imitation as maximum-entropy RL with a KL-based adherence term and dual optimization. The main objective is

PRESERVED_PLACEHOLDER_^^^^7query77^^^^

and its practical Lagrangian, implemented atop SAC, combines reward, entropy, and a demonstration-based term derived from PRESERVED_PLACEHOLDER_^^^^7query7^^^^8. The method is model-free, off-policy, and hybrid offline–online: it uses a fixed demonstration set PRESERVED_PLACEHOLDER_^^^^7query7^^^^9 alongside environment interaction through a replay buffer PRESERVED_PLACEHOLDER_^^^^7ti:\7query7^^^^. It does not project actions, does not construct an explicit safe action set, and does not guarantee per-step feasibility; its constraint signal is transferred from demonstration occupancy patterns rather than hard action masks (&&&^^^^7max_results7^^^^&&&).

Formulation	Core mechanism	Test-time handling
DTWIL	Feasible surrogate demonstrations via MPC + DTW	Standard IL on surrogate data
Constraint-aware Behavior Cloning	BC plus learned one-step safety penalty	Only policy network used
SCOPIL	Max-ent RL plus KL imitation and dual descent	No mask or projection layer

This comparison clarifies an important terminological issue. In the strict sense, ACIL refers to the explicit action-constrained setting formalized through PRESERVED_PLACEHOLDER_^^^^7ti:\7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ and addressed by DTWIL. In a broader and practically common sense, the term also overlaps with safety-aware or constraint-aware imitation methods that shape policies through predicted-state feasibility or demonstration-induced constraint information rather than online action projection.

^^^^7 OR ti:\7^^^^. Empirical findings

The strict ACIL paper evaluates DTWIL on four continuous-control domains: Maze^^^^7max_results7^^^^d-Medium-v^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^, HalfCheetah-v^^^^7query7^^^^, Hopper-v^^^^7max_results7^^^^, and Robosuite Table-Wiping. The comparisons include both learning-from-demonstration and learning-from-observation baselines: GAIL, BCO, GAIfO, OPOLO, CFIL-s, CFIL-sa, SAIL, and DIFO. All online methods are limited to ^^^^7 OR ti:\7query7^^^^K environment steps during training. Across nearly all tasks, DTWIL is reported to perform best or among the best in return while also achieving the smallest DTW distance to expert trajectories. Representative results include Maze^^^^7max_results7^^^^D M+B return PRESERVED_PLACEHOLDER_^^^^7ti:\7max_results7^^^^ with DTW distance PRESERVED_PLACEHOLDER_^^^^7ti:\7query7^^^^; HalfCheetah HC+B return PRESERVED_PLACEHOLDER_^^^^7ti:\7ti:\7^^^^, compared with PRESERVED_PLACEHOLDER_^^^^7ti:\7 OR ti:\7^^^^ for SAIL, and DTW distance PRESERVED_PLACEHOLDER_^^^^7ti:\7 OR ti:\7^^^^ versus PRESERVED_PLACEHOLDER_^^^^7ti:\77^^^^ for SAIL; Hopper H+B return PRESERVED_PLACEHOLDER_^^^^7ti:\7^^^^8; and Table-Wiping W+L^^^^7max_results7^^^^ return PRESERVED_PLACEHOLDER_^^^^7ti:\7^^^^9. Ablations show that DTW-based alignment outperforms PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7query7^^^^ alignment on HC+B (PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ versus PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7max_results7^^^^) and H+B (PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7query7^^^^ versus PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7ti:\7^^^^), that asynchronous progression dominates synchronous progression on Hopper (PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7 OR ti:\7^^^^ versus PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7 OR ti:\7^^^^), that ERC markedly improves Hopper performance (Return-BC PRESERVED_PLACEHOLDER_^^^^7 OR ti:\77^^^^ with ERC versus PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7^^^^8 without), and that replanning every step is superior to executing ^^^^7query7^^^^ or ^^^^7 OR ti:\7^^^^ open-loop steps (&&&^^^^7query7^^^^&&&).

Constraint-aware Behavior Cloning is validated in CARLA-based autonomous racing with customized vehicle dynamics under three settings: image-feedback path following at conservative speed, full-state autonomous racing, and image-feedback autonomous racing. The reported metrics include successful consecutive laps completed without constraint violation, training epochs required to reach sustained success, lap time, lap-time variability, and qualitative distance from walls. In the most difficult image-feedback racing setting, the learned policy completed ^^^^7 OR ti:\7query7^^^^ consecutive laps within 8^^^^7query7^^^^ epochs, while the baseline failed to exceed ^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7query7^^^^ laps. Reported early-stopped rollout statistics were average lap time PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7^^^^9 s, max lap time PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7query7^^^^ s, min lap time PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ s, and expert average lap time PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7max_results7^^^^ s. In full-state racing, the method achieved higher return with less training effort, similar average lap time to the baseline and expert, and lower lap-time variability, which the paper interprets as greater consistency (&&&^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^&&&).

SCOPIL is evaluated in the Marble Maze environment with 9 discrete board-rotation actions under three settings: Simple, Multiple-constraints, and Two-modes. In the Simple setting, SCOPIL reports horizontal violations PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7query7^^^^, cyclic violations PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7ti:\7^^^^, and total violations PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7 OR ti:\7^^^^, versus total violations PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7 OR ti:\7^^^^ for ICRL and PRESERVED_PLACEHOLDER_^^^^7 OR ti:\77^^^^ for SAC. In the Two-modes setting, it reports horizontal violations PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7^^^^8, cyclic violations PRESERVED_PLACEHOLDER_^^^^7 OR ti:\7^^^^9, and total violations $$\pi^*:=\underset{\pi\in \Pi_{\mathcal{C}}}{\arg\min}\; \mathbb{E}_{\tau_e\sim \pi_e,\;\tau\sim \pi} \big[d(\sigma(\tau), \sigma(\tau_e)) \mid s_0\sim \mu\big].$$^^^^7query7^^^^. The paper emphasizes that safety is not encoded in the reward, which remains task-related, and attributes its multimodal robustness to the demonstration-based constraint signal and the use of $$\pi^*:=\underset{\pi\in \Pi_{\mathcal{C}}}{\arg\min}\; \mathbb{E}_{\tau_e\sim \pi_e,\;\tau\sim \pi} \big[d(\sigma(\tau), \sigma(\tau_e)) \mid s_0\sim \mu\big].$$^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^ rather than adversarial occupancy matching (&&&^^^^7max_results7^^^^&&&).

Taken together, these results support three distinct empirical claims. First, strict ACIL benefits from replacing infeasible demonstrations with realizable surrogates rather than attempting direct expert matching. Second, training-time safety shaping can materially improve consistency and constraint satisfaction even when only the policy is deployed at test time. Third, demonstration-induced constraint information can reduce violations and preserve multimodal safe behavior without explicit constraint costs.

^^^^7 OR ti:\7^^^^. Limitations, misconceptions, and conceptual boundaries

A recurrent misconception is that ACIL is equivalent to clipping or projecting actions into a feasible set. The ACIL formulation rejects that reduction: projection changes the transition and thereby alters future occupancy, so it does not solve the imitation problem by itself (&&&^^^^7query7^^^^&&&). A second misconception is that state-only imitation automatically resolves action infeasibility. The ACIL analysis argues otherwise, because state-only learners may still target trajectories that were generated by unconstrained actions and remain physically unrealizable for the constrained imitator (&&&^^^^7query7^^^^&&&).

The main limitations differ across formulations. DTWIL relies on a learned forward model, incurs substantial offline computation from MPC, CEM, and DTW scoring, and assumes the expert and imitator share the same state and action dimensions. The paper does not provide strong formal optimality guarantees and notes possible failure modes when constraints are extremely tight, when the expert path remains infeasible even after temporal warping, when instability amplifies model error, or when the planning horizon is too short (&&&^^^^7query7^^^^&&&). Constraint-aware Behavior Cloning explicitly provides no formal guarantees of constraint satisfaction for the deployed policy, because safety is enforced only as a soft penalty via learned dynamics, learned safe-set probabilities, heuristic safety labels, and no runtime filter (&&&^^^^7Action-Constrained Imitation Learning arXiv (Yeh et al., 20 Aug 2025) constraint-aware imitation learning autonomous racing (Cao et al., 10 Mar 2025) SCOPIL (Papadopoulos et al., 9 Jul 2025)7^^^^&&&). SCOPIL likewise offers no hard guarantees of zero violations at execution time, since it does not learn an explicit admissible action set and enforces constraints only through a dual-weighted objective shaped by demonstrations (&&&^^^^7max_results7^^^^&&&).

The conceptual boundary of ACIL is therefore best drawn along the mechanism of feasibility. In the narrow sense, ACIL studies imitation when the learner’s feasible action set is smaller than the expert’s and standard occupancy matching is structurally unattainable. In the broader neighboring literature, feasibility is often encoded indirectly—through predicted successor-state safety, backward-reachability surrogates, or demonstration-induced occupancy constraints. This suggests that contemporary work on ACIL is organized around a common question—how to imitate when expert behavior is not directly executable—while differing sharply in whether feasibility is enforced by surrogate demonstration generation, by training-time safety shaping, or by soft distributional adherence.