{"id":"d96dad2e-95e7-4bec-b32e-6cc849d0acb3","arxiv_id":"2506.22894","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"An RL-based drift motion planner with a predictive safety filter outperforms an MPC benchmark on tracking error and compute in simulation.","lead":"This paper presents a reinforcement learning planner for autonomous drifting that learns the road curvature and friction online, with a predictive safety filter constraining the curvature before a model-predictive drift controller tracks it. In Matlab-Carsim simulation, the safe planner cuts mean lateral deviation roughly by half and mean computation time by 72% compared with an MPC baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The PSF's safety guarantee rests on a frozen-state prediction model (17); Fig. 5a's early termination shows the filter does not actually enforce |e|<eth. The 'Safe RL' claim is therefore not established.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I would raise: the PSF's safety constraints are enforced on a simplified model with frozen V, beta, and r, which is not valid during aggressive drifting. The paper's own Fig. 5a provides empirical evidence that the filter does not prevent large lateral errors, as the episode terminates at |e|>emax even with the PSF active. This directly undermines the 'Safe RL' label and contribution 2. I considered whether the statistical robustness of the performance comparison or the fairness of the MPC baseline is more load-bearing, but those concerns affect the strength of the performance claim rather than the core safety claim, and the paper's title and abstract emphasize safety. Since the reader already assigned CONDITIONAL with this concern, my read does not change the verdict. The concrete test above would settle whether the concern actually lands by measuring constraint violations on the real trajectory and by testing a full-model PSF as a counterfactual.","tokens_in":12046,"tokens_out":6163,"duration_ms":68938,"concrete_test":"Re-run the trained Safe RL policy on the training track while logging, at each RL step, the PSF's predicted trajectory over the Ns=10 horizon and the actual Carsim trajectory; compute the violation rate of |e|<eth and |Delta psi|<Delta psi_th along the actual trajectory. If any violation occurs, the PSF does not enforce safety on the real dynamics. A stronger version: replace model (17) in the PSF with the full nonlinear model (1)-(3) and compare violation counts; if violations disappear, the frozen-state assumption is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is a Safe RL planner, so the safety filter's guarantee is load-bearing. The PSF solves OCP (16) with prediction model (17), in which V, beta, and r are held constant over the horizon (footnote 1). The constraints |e|<eth and |Delta psi|<Delta psi_th are imposed on this simplified model, not on the full drift dynamics (1)-(3) or the actual Carsim vehicle. During aggressive drifting, V, beta, and r change substantially over the 10-step horizon, so the PSF's predicted trajectory can diverge from the realized trajectory. This is not only a theoretical gap: Fig. 5a shows an Episode-500 run where the PSF is active and the vehicle still triggers early termination at |e|>emax=5 m, far beyond eth=1.5 m. Thus the filter does not, in practice, prevent unsafe states, and the claim that the proposed planner 'ensures safe and efficient learning' is unsupported. Absent a proof that model (17) is a valid control barrier or that constraint satisfaction transfers to the full dynamics, the label 'Safe' and contribution 2 are not established. The performance advantage over the MPC benchmark may survive, but the central claim as stated concerns a safe planner, so this is a load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical Safe Reinforcement Learning (RL) motion planner for autonomous drifting. An RL agent (DDPG) learns a curvature offset and a friction coefficient correction to feed a Model Predictive Drift Controller (MPDC). A Predictive Safety Filter (PSF) solves a short-horizon optimal control problem to adjust the RL curvature reference so that a simplified path-tracking model satisfies |e|<eth and |Δψ|<Δψth. The authors validate in Matlab-Carsim simulations, reporting that the safe RL planner outperforms an MPC benchmark in tracking accuracy, drift-state RMSE, and computation time, and that the PSF accelerates training convergence. The paper also presents ablations isolating the effects of learning κ, learning μ, and adding the PSF, plus generalization tests on three unseen tracks.","tokens_in":12384,"tokens_out":5264,"duration_ms":56393,"significance":"If the safety and performance claims hold, this is a useful contribution to learning-based motion planning for highly dynamic maneuvers: it combines model-free RL with a modular model-based safety filter, avoiding expert demonstrations while keeping a principled controller. The paper is careful in its ablation design: the low-level MPDC is unchanged across all comparisons, and the reported improvements are consistent across lateral error, heading error, drift-state RMSE, computation time, and generalization tracks. The computational-efficiency comparison is a notable strength, as is the explicit training-time cost of the PSF. However, the central 'safe' claim is currently stated too strongly relative to the evidence, so the significance hinges on whether the safety guarantee can be substantiated for the full closed-loop system.","major_comments":[{"comment":"The safety guarantee of the PSF is not established. The prediction model (17) keeps V, beta, and r constant over the horizon, and the constraints |e|<eth and |Δψ|<Δψth are enforced on this simplified kinematic model, not on the full drift model (1)-(3) or the Carsim vehicle. Since the low-level MPDC tracks a drift equilibrium rather than directly imposing path-tracking constraints, the realized closed-loop errors can differ substantially from the PSF predictions. This is not merely a theoretical gap: Fig. 5a shows an Episode-500 run where the PSF is active (κ0 deviates from κRL) yet the vehicle still triggers early termination at |e|>emax=5 m, well above eth=1.5 m. Thus the statements that the PSF 'prevents unsafe states' (contribution 2) and that the method 'ensures safe and efficient learning' (abstract) are unsupported. The authors should either provide a formal safety argument for the full closed-loop system (e.g., robust invariance or tube-based reasoning) or present a thorough empirical safety evaluation over all training episodes and test tracks, and revise the wording to 'reduces safety violations' rather than 'ensures safety'.","section":"Section V, Eq. (17), footnote 1, Fig. 5a"},{"comment":"The PSF OCP formulation does not discuss recursive feasibility or the behavior when the OCP is infeasible. The RL agent can propose a curvature for which no solution of (16) satisfies the constraints; the paper does not specify what action is executed in that case, how the RL update handles the lack of a safe action, or whether the MPDC can still receive a feasible reference. Since the PSF is the core safety mechanism, the absence of a feasibility analysis leaves a gap in the safety claim and in the practical implementation of the training loop.","section":"Section V, Eq. (16)"}],"minor_comments":[{"comment":"The computational load in Fig. 3c is reported without clarifying whether it includes the MPDC solve time. Given that the RL planner operates at a 0.1 s sampling time and the reported Safe RL mean load is 0.137 s, the reader cannot determine whether the full system meets real-time constraints; please state explicitly what the box-plot values include and add the aggregate per-step latency.","section":"Section VI-H, Fig. 3c"},{"comment":"The claim of 'no prior expert knowledge' should be qualified: the equilibrium calculation in Section III-B fixes δ0 = -0.5 rad and uses a model-based drift dynamics, which constitutes model knowledge. Clarify that the claim refers to the absence of expert demonstrations or hand-tuned initial policies, not to the absence of any model-based prior.","section":"Section I-B, contribution 3"},{"comment":"In the reward description, the phrasing 'penalizes further the lateral error e further' is redundant; consider rewording to 'additionally penalizes the lateral error e once it exceeds the threshold'.","section":"Section IV-B-3"},{"comment":"When reporting percentage improvements over the MPC benchmark, please include the absolute values of the errors in the same sentence (they appear only in Fig. 3), so the reader can assess the practical magnitude of the gains.","section":"Section VI-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the performance results appear well-consistently presented. The main issue is not the performance comparison but the strength of the safety claim: the frozen-state model (17) and the evidence in Fig. 5a directly undermine the 'ensures safety' wording. This is fixable within the manuscript's scope either by adding formal closed-loop guarantees or by tempering the claims and adding a rigorous empirical safety analysis. I recommend major revision rather than rejection, because the underlying architecture and the comparative evaluation are valuable even if the safety guarantee is ultimately stated as heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the integration of RL-based online curvature and friction adaptation with a predictive safety filter is a real and sensible step for autonomous drifting, and the simulations support the performance claim against the MPC baseline. But the word 'Safe' is load-bearing, and the filter's guarantee is not actually shown: model (17) freezes V, beta, and r, and Fig. 5a shows an early termination at |e|>emax with the filter active. So the safety contribution is a heuristic that reduces errors, not a guarantee.\n\nWhat is genuinely new: I don't recall another drifting planner that uses RL to adjust both the reference curvature and the friction coefficient in closed loop with a safety filter. The ablation separating curvature learning from friction learning is good, and the head-to-head against the enhanced MPC from [15] is fair. The 3.6x computation advantage is credible, and the training curves show the PSF does help convergence.\n\nSoft spots: the central one is the mismatch between the 'Safe RL' language and what the PSF actually does. Footnote 1 admits V, beta, and r are held constant over the 10-step horizon. During hard drifting those states change fast, so the OCP's constraints |e|<1.5 m and |Δψ|<0.2 rad do not transfer to the real vehicle. Fig. 5a is not just a theoretical worry: the filter is active and the episode still terminates because |e| exceeds 5 m. That means the claim to 'prevent unsafe states' is overstated. The right fix is to soften the guarantee language or add an analysis of when constraint satisfaction on (17) implies safety on (1)-(3), e.g., Lipschitz bounds on the drift states over the horizon.\n\nAdditional minor issues: single training seed, no error bars on the box plots, and the generalization test with only three tracks is a start but not convincing. No code or data is available, which lowers confidence.\n\nBottom line: a useful engineering contribution that deserves a serious referee, but not with the current safety claims. I'd send it to review with major revisions expected: soften the claims, add formal or empirical evidence about the filter's effect on true constraints, and ideally release code/data. I wouldn't cite it as a safe RL method in my own work, but I might cite the empirical comparison.","headline":"A useful RL+PSF drifting planner with real performance gains, but the 'Safe' claim outruns the evidence: the filter's frozen-state model does not enforce the stated constraints, and Fig. 5a shows a violation.","tokens_in":12882,"tokens_out":3457,"would_cite":false,"duration_ms":31069,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A predictive safety filter lets a reinforcement-learned planner drift safely and beat an MPC benchmark.","keywords":["safe reinforcement learning","predictive safety filter","autonomous drifting","motion planning","drift equilibrium","road friction learning","path tracking","DDPG"],"falsifier":"Run the trained Safe RL planner on a high-fidelity or physical vehicle through a transient drift where yaw rate and sideslip change rapidly, and record whether the actual lateral error exceeds $e_{\\mathrm{th}} = 1.5$ m even though the PSF's filtered curvature satisfied $|e|<e_{\\mathrm{th}}$ in its prediction model. One such violation during normal PSF operation would refute the claim that the filter prevents unsafe states.","tokens_in":11876,"feed_emoji":"🏎️","tokens_out":5037,"duration_ms":54587,"temperature":0.7,"pith_summary":"This paper tries to establish that a reinforcement-learning motion planner can plan autonomous drifting along a general path without expert demonstrations, provided a predictive safety filter screens its curvature commands. The planner learns two quantities online: a correction to the reference curvature and a correction to the road friction coefficient, which improve the drift-equilibrium calculation of a model predictive drift controller. If correct, the approach offers a learning-based alternative to MPC-based drift planning that is safer during training and faster online. The claim is supported by simulations where the safe RL planner reduces mean lateral deviation by about half and mean heading error by about 60 percent relative to the benchmark, while using 3.6 times less computation.","feed_headline":"Safety-filtered RL beats MPC for autonomous drifting","feed_subtitle":"Learned curvature and friction corrections keep the drift path within bounds while running 3.6x faster.","key_machinery":"The central object is the Predictive Safety Filter (PSF), an optimal control problem that minimizes the deviation between the RL-planned curvature and the filtered curvature subject to the path-tracking model $\\dot e = V\\sin\\Delta\\psi$ and $\\Delta\\dot\\psi = \\dot\\beta + r - \\kappa V\\cos\\Delta\\psi/(1-\\kappa e)$, with constraints $|e|<e_{\\mathrm{th}}$ and $|\\Delta\\psi|<\\Delta\\psi_{\\mathrm{th}}$ enforced over a short prediction horizon. The RL agent, trained with DDPG, simultaneously learns curvature adjustment $\\kappa_{\\mathrm{RL}} = \\kappa_r + \\varepsilon_\\kappa$ and friction adjustment $\\mu_{\\mathrm{RL}} = \\mu_n + \\varepsilon_\\mu$; these corrections refine the drift-equilibrium calculation that the low-level Model Predictive Drift Controller tracks. The PSF sits between the RL agent and the controller, rejecting or modifying curvature commands that would violate the safety constraints.","core_discovery":"On the paper's own terms, the proposed Safe RL planner couples a DDPG agent that outputs reference-curvature and road-friction corrections with a Predictive Safety Filter that re-solves a short-horizon optimal control problem at every step to keep lateral and yaw errors within track bounds, and then feeds the filtered curvature to a model predictive drift controller. In Matlab-Carsim simulations over an 18-second clothoid path, this combination produced 49.6% lower mean lateral deviation and 61.1% lower mean heading error than an enhanced MPC planner benchmark, while requiring 3.6 times less average computation time. The same architecture also generalizes to three unseen tracks with varying curvature, keeping lower tracking RMSE than the fixed-parameter MPC benchmark on every test track.","pith_inferences":["The safety guarantee is only as strong as the PSF's frozen-parameter prediction model: if velocity, sideslip, and yaw rate change substantially within the prediction horizon during aggressive transients, the filtered curvature could violate the true constraints even though the simplified model reports a safe trajectory.","The same PSF-plus-RL architecture could transfer to other underactuated, safety-critical maneuvers such as emergency obstacle avoidance on low-friction roads, where the RL agent adapts to unknown parameters and the filter rejects unsafe reference commands.","A testable extension is to replace the frozen-parameter PSF model with a reachable-set or robust-tube forecast; the resulting losses in tracking performance and gains in safety could be measured directly against the present simulation results.","Because PSF corrections diminish as training progresses, the filter could be omitted at inference time to save computation, but doing so would sacrifice the explicit safety layer that the paper highlights as its contribution."],"forward_implications":["A learning-based drift planner can run online in real time on modest hardware, whereas the MPC benchmark's average 0.495 s solve time is far too slow for the 0.05 s control loop.","The predictive safety filter accelerates training convergence and raises average reward, so safety filtering acts as a training-time regularizer as well as a deployment safeguard.","Learning the road friction coefficient online reduces drift-state RMSE by 11.5% relative to a curvature-only RL planner, showing that model-error correction matters for drift performance.","The safe RL planner retains its tracking advantage over the MPC benchmark on unseen test tracks without retuning, supporting the paper's claim of generalization to variable-curvature environments."],"supporting_citations":[{"why":"Serves as the MPC planner benchmark, enhanced with variable-curvature path prediction, that the safe RL planner must beat.","marker":"[15]"},{"why":"Supplies the predictive safety filter formulation that the paper adapts to filter the RL curvature reference.","marker":"[32]"},{"why":"Provides the vehicle path-tracking dynamics model used inside the PSF prediction equation.","marker":"[28]"},{"why":"Provides the single-track drift vehicle model with weight transfer and wheel speed used by the MPDC.","marker":"[5]"},{"why":"Provides the Pacejka tire model that links tire forces to slip angles and friction coefficient.","marker":"[27]"},{"why":"Establishes the drift equilibrium point calculation used to derive desired drift states.","marker":"[29]"},{"why":"Defines the DDPG algorithm used as the RL agent for curvature and friction learning.","marker":"[31]"}],"fun_headline_variants":["Safe RL drifts better than MPC: 49% less lateral error","RL with safety filter cuts drift tracking error by 61%","Safety-filtered RL drifts with 3.6x less compute","Predictive safety filter enables safe RL drifting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The safety filter checks the RL curvature using a simplified forecast that freezes the vehicle's speed, sideslip, and yaw rate, so its \"safe\" output may not be safe for the real vehicle during aggressive drifting.","fun_headline_variants_meta":{"raw":{"variants":["Safe RL drifts better than MPC: 49% less lateral error","RL with safety filter cuts drift tracking error by 61%","Safety-filtered RL drifts with 3.6x less compute","Predictive safety filter enables safe RL drifting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000385,"raw_usage":{"total_tokens":2004,"prompt_tokens":885,"completion_tokens":1119,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1048}},"tokens_in":501,"tokens_out":1119,"duration_ms":10128,"temperature":1.0,"reasoning_tokens":1048,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:54:39.985393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained Safe RL planner on a high-fidelity or physical vehicle through a transient drift where yaw rate and sideslip change rapidly, and record whether the actual lateral error exceeds $e_{\\mathrm{th}} = 1.5$ m even though the PSF's filtered curvature satisfied $|e|<e_{\\mathrm{th}}$ in its prediction model. One such violation during normal PSF operation would refute the claim that the filter prevents unsafe states.","supporting_citations":[{"cited_title":"A novel model predictive controller for the drifting vehicle to track a circular trajectory,","cited_arxiv_id":null,"evidence_quote":"Serves as the MPC planner benchmark, enhanced with variable-curvature path prediction, that the safe RL planner must beat."},{"cited_title":"A predictive safety filter for learning-based racing control,","cited_arxiv_id":null,"evidence_quote":"Supplies the predictive safety filter formulation that the paper adapts to filter the RL curvature reference."},{"cited_title":"Toward automated vehicle control beyond the stability limits: Drifting along a general path,","cited_arxiv_id":null,"evidence_quote":"Provides the vehicle path-tracking dynamics model used inside the PSF prediction equation."},{"cited_title":"Modeling and control for dynamic drifting trajectories,","cited_arxiv_id":null,"evidence_quote":"Provides the single-track drift vehicle model with weight transfer and wheel speed used by the MPDC."},{"cited_title":"Tyre modelling for use in vehicle dynamics studies,","cited_arxiv_id":null,"evidence_quote":"Provides the Pacejka tire model that links tire forces to slip angles and friction coefficient."},{"cited_title":"A controller framework for autonomous drifting: Design, stability, and experimental validation,","cited_arxiv_id":null,"evidence_quote":"Establishes the drift equilibrium point calculation used to derive desired drift states."}],"review_version":1}