Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Learned blocking AI delays overtakes from 8.8 to 14.6 seconds in simulated racing

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 08:49 UTC pith:SQ2ORYD5

load-bearing objection The architecture is a sensible new twist on RL-MPC, but the headline 8.8→14.6 s overtake delay is a segment-subset artifact and the full-track gain is materially smaller; still worth refereeing with revisions. the 3 major comments →

arxiv 2607.20973 v1 pith:SQ2ORYD5 submitted 2026-07-23 cs.RO

Deep Reinforcement-Learning-Guided Model Predictive Control for Preventing Overtakes in Autonomous Racing

Yufei Xi , Yijie Liao , Tulga Ersal This is my paper
classification cs.RO
keywords autonomous racingdefensive blockingmodel predictive controldeep reinforcement learningSoft Actor-CriticFrenet framespatial envelopefriction-limited dynamics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that defensive driving in autonomous racing can be framed not as lap-time minimization but as spatial occupancy regulation: the ego vehicle learns where to position itself on the track to deny a faster opponent a passing corridor. It proposes a hierarchical control stack in which a Soft Actor-Critic policy operates in the Frenet frame, generating spatial references that are embedded into a nonlinear model predictive controller as regularization terms. In simulation on a held-out real track, this RL-guided MPC increases average overtake time from 8.8 s to 14.6 s and cuts opponent progress by roughly 33 m, while using 83.4% of available tire force and solving in 33.3 ms on average. The sympathetic reading is that high-level blocking intent can be realized in dynamically feasible, real-time-controlled trajectories, and that the gains come from positioning rather than from exceeding physical limits.

Core claim

On its own terms, the paper claims that defensive blocking can be decomposed into a learned spatial-referencing problem and a constraint-satisfying tracking problem. An RL policy trained with Soft Actor-Critic outputs future lateral offsets and arc-length adjustments in the Frenet frame; these are transformed to Cartesian waypoints and added as a reference-tracking term in a Spatial Envelope MPC that enforces the nonlinear single-track dynamics, friction-circle constraints, and actuator limits. In closed-loop adversarial simulation against a progress-maximizing opponent with a 1.0 s preview and elevated tire friction, this framework delays overtaking from an average of 8.8 s (baseline MPC wi

What carries the argument

The central mechanism is the spatial reference set R = {(s_i^π, d_i^π, w_i)} generated by a Soft Actor-Critic policy in the Frenet frame. These learned geometric references are embedded into the nonlinear MPC cost as a collocation-aligned reference-tracking term J_ref, which continually attracts the optimal trajectory toward the RL's defensive intent while the MPC enforces dynamic feasibility. This 'policy-induced regularization' is what bridges high-level adversarial reasoning and low-level constraint-satisfying control, and it is the component ablated in the baseline comparison.

Load-bearing premise

The only adversary tested is a progress-maximizing MPC with a 1.0-second preview of the ego's predicted motion and elevated tire friction; if the opponent anticipated the ego's blocking (e.g., via game-theoretic reasoning or adaptive learning), the overtake-time gains might shrink or reverse.

What would settle it

Run the same RL-guided MPC controller against a game-theoretic opponent (or one with a longer preview / learned blocking-aware policy) on the same track segments; if the average overtake time falls back to near the 8.8 s baseline, the defensive gains are an artifact of the specific opponent model rather than a general property of the framework.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Defensive driving can be formulated as a learned spatial-objective problem, offering an alternative to explicit game-theoretic reasoning in multi-agent racing.
  • The framework runs in real time (mean 33.3 ms solve on the tested hardware), suggesting it could be deployed on embedded platforms for closed-loop adversarial interaction.
  • Tire-force utilization stays near friction limits under both controllers, indicating the overtake-time improvement stems from strategic positioning rather than artificially exceeding dynamic limits.
  • Frenet-domain policy learning transfers to a held-out real track (Thunderhill West), supporting the claim that the approach generalizes across track geometry.
  • The overtake-time definition (time until opponent progress exceeds ego by a threshold) provides a concrete, measurable metric for evaluating defensive performance in racing scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The measured 5.8 s delay is likely opponent-model-dependent: a game-theoretic opponent that anticipates blocking, or even one with a longer preview, could erode or reverse the gain; a natural stress test is to evaluate against adaptive or strategic adversaries.
  • The paper's observation that RL places some dynamically infeasible side-by-side references, which the MPC then interprets as speed reduction and lateral movement, suggests the hierarchy may tolerate imperfect reference generation—an implicit robustness property worth probing explicitly.
  • The common defensive tactic of deliberately slowing entry speed to force opponent deceleration hints at a transferable principle: denying an opponent's acceleration advantage can be more effective than merely holding a racing line.
  • Because only one held-out track and one opponent policy were tested, the reported averages should be read as an upper bound on expected gains until multi-track, multi-opponent evaluations are performed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a hierarchical framework for defensive blocking in autonomous racing. A Soft Actor-Critic (SAC) policy operates in a Frenet frame and generates spatial reference points, which are embedded as a regularization term in a nonlinear Spatial Envelope MPC that enforces dynamic feasibility and friction constraints. The opponent is a progress-maximizing MPC with a 1.0 s preview of the ego's predicted motion and elevated friction. Evaluation on the Thunderhill West circuit in simulation reports, over Segments 5–9, an increase in average overtake time from 8.8 s to 14.6 s, reduced opponent distance, 83.4% tire utilization, and 33.3 ms mean solve time. The paper claims the framework enables real-time defensive control in closed-loop adversarial interaction.

Significance. If the quantitative claims hold, the idea of using RL-generated spatial references as policy-induced regularization inside a nonlinear MPC is a reasonable and potentially useful way to inject strategic intent while retaining dynamic feasibility. The held-out track evaluation is a strength, and the reported solve times support real-time feasibility. However, the evidence as presented is conditional: the headline average is computed on a post-hoc subset of segments, there are no repeated trials or error bars, and the opponent model is fixed. The paper's limitations section is candid, but the Abstract and Section IV.B state the result more strongly than the reported data justify. The contribution is better framed as a demonstration of a specific RL-MPC defensive mechanism than as a general claim about delaying faster opponents.

major comments (3)
  1. [§IV.B, Table II] The abstract's headline '8.8 s to 14.6 s' is the average over Segments 5–9 only, selected because the baseline already prevents overtaking in Segments 1–4. Using the paper's own convention that a non-overtake is set to the 20 s horizon (§III.A), Table II implies a full-track baseline average of (4×20+14.2+7.6+6.0+7.2+8.8)/9 ≈ 13.8 s and a proposed average of (4×20+15.4+20.0+13.8+17.6+6.3)/9 ≈ 17.0 s, i.e. about 3.3 s (24%) rather than 5.8 s (66%). Moreover, Segment 9 reverses the trend (6.3 s vs 8.8 s). The central quantitative claim as stated is not supported. Please report all nine segments, state the selection rule explicitly, and either pre-specify the subset or analyze the full track.
  2. [§IV.A–IV.B, Table II] No repeated trials, random seeds, or varied initial conditions are reported; each segment appears as one run per controller. Overtake time is a stochastic function of the initial relative state and the warm-started solver, and the paper provides no error bars, confidence intervals, or paired analysis. Given that Segment 9 shows the opposite sign, claims of 'significantly increases' and 'significantly reducing' are not statistically supported. Please add multiple runs per segment and report the dispersion, or soften the language accordingly.
  3. [§IV.A] The opponent is a single, deterministic progress-maximizing MPC with a 1.0 s preview of the ego's predicted motion and elevated friction. The paper's central claim concerns delaying a 'faster opponent' (Abstract), but only this one opponent model is tested. If the opponent anticipated the ego's blocking (e.g., game-theoretic or adaptive policy), or if the preview horizon or friction advantage were changed, the overtake-time gains could shrink or reverse. The limitations section mentions this, but the evaluation should include at least a sensitivity study over opponent horizons/models, or the claims should be explicitly restricted to this opponent class.
minor comments (5)
  1. [§III.A, Table I] The overtake definition uses thresholds Δs_pass and t_hold, but their numerical values are not given in Table I or in the text. Please add them.
  2. [§IV.B] The decision to exclude Segments 1–4 is described qualitatively ('the baseline MPC no longer induces such implicit blocking'). Please provide quantitative evidence for the claimed implicit blocking, e.g., lateral gap or passing-corridor analyses, so the reader can assess the selection.
  3. [§III.D, Eq. (15)] The notation ∥·∥_{Q_s} and ∥·∥_{Q_v} is introduced without defining the matrices' dimensions or the norm type. Please specify.
  4. [§IV.C] There is a duplicated sentence: 'It interprets such reference points as a strategic decision to reduce speed and move laterally.' appears twice in the qualitative analysis. Also, the time-stamp labels in Fig. 6 are very small and hard to read.
  5. [§V] The conclusion says 'delays its opponent by a total of 5.8 s', but the reported 5.8 s is the average difference over Segments 5–9. Please clarify whether this is a per-segment average or a cumulative total.

Circularity Check

0 steps flagged

No derivation-level circularity; the only self-citation is a deliberately non-load-bearing ablation baseline, and the reward/evaluation alignment is not an equation-level identity.

full rationale

The paper's central claim is an empirical benchmark result, not a theorem derived from first principles. The proposed controller generates spatial references R (Eq. 6) that enter the MPC as D_ref (Eq. 15), while the baseline is the same Spatial Envelope MPC 'with D_ref removed' (Section III.E). The 8.8 s to 14.6 s overtake-time increase is therefore measured against an explicit internal ablation, not predicted from the reward function by construction. The closest alignment is between the defensive reward term F_def (Eq. 13), which rewards positive Frenet lead Δs_rel, and the overtake-time metric (Section III.A), which is the first time the opponent's progress exceeds the ego's by a threshold. This is a reward/evaluation alignment common in empirical RL studies, not a self-definitional reduction: F_def is a weighted, thresholded reward used during training on procedurally generated tracks, while OT is a threshold-crossing time over a 20 s horizon on the held-out Thunderhill West track, and the MPC constraints can fail to realize the learned references (as the paper itself acknowledges for Segment 9). No fitted parameter is renamed as a prediction. The baseline citation to 'Spatial Envelope MPC [1]' is a potential self-citation, but it is deliberately used only as an ablation and is not invoked as evidence for the central claim, so it is not load-bearing. The paper explicitly flags its limitations in Section V (simulation-only, one held-out track, single opponent policy) and the Segments 5-9 evaluation choice is a validity/selective-reporting concern rather than circularity. Overall: no significant circularity; score 2 reflects only the minor non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced; the 'spatial reference set' is a control construct, not an independent entity.

free parameters (5)
  • Reward weights λ1..λ4 = not reported (fixed)
    Equation (11) defines reward as a weighted sum; weights are chosen by hand and shape the learned policy; no sensitivity analysis.
  • MPC weights λ_ref, λ_safe, λ_smooth, Q, w_tube, w_avoid = Table I values (e.g., Q=diag(0.01,0.01), w_tube=10, w_avoid=1500)
    Balance reference tracking, safety, and smoothness; hand-tuned; affect the trade-off between defense and safety.
  • Overtake definition thresholds Δs_pass and t_hold = not reported
    Definition of overtake time in Section III.A depends on these thresholds; results are sensitive to them.
  • Reference node weights w_i = fixed (not learned)
    Weights for each spatial reference point in Eq. (6); influence how strongly each future node is tracked.
  • RL training hyperparameters = lr=3e-4, batch=256, buffer=1e6
    Table I; chosen for stable SAC training; not swept.
axioms (4)
  • domain assumption The single-track bicycle model with Fiala tire forces is an adequate representation of vehicle dynamics for this task.
    Used as MPC prediction model and simulation environment; inaccuracies in tire modeling would change the friction-circle feasibility and the measured 83.4% utilization.
  • domain assumption The opponent vehicle's strategy is a meaningful proxy for a faster adversary.
    The opponent optimizes progress with 1.0s ego preview and higher friction coefficients; the defense results depend on this behavior being representative.
  • ad hoc to paper The fixed-weight reward decomposition captures the intended defensive objective.
    Eq. (11) combines progress, lead, collision, and boundary penalties with manually set λ_i; if weights are poorly chosen, learned references may not maximize overtake delay.
  • domain assumption Embedding learned references as an MPC regularization term transfers strategic intent without violating constraints.
    Section III.D assumes that a reference-tracking cost will make the MPC 'interpret' infeasible references as speed/lateral adjustments rather than causing instability.

pith-pipeline@v1.3.0-alltime-deepseek · 13074 in / 11269 out tokens · 112395 ms · 2026-08-01T08:49:12.317129+00:00 · methodology

0 comments
read the original abstract

This paper addresses defensive blocking in autonomous racing, where a vehicle must prevent a faster opponent from overtaking while operating near its dynamic limits. Different from lap-time minimization, we formulate defense as a spatial occupancy regulation problem via a hierarchical reinforcement-learning guided model predictive control framework. A Soft Actor-Critic strategic layer operates in the Frenet domain to generate geometry-aware defensive references, which are embedded into the nonlinear model predictive control formulation as spatial regularization under friction constraints. Evaluated on the Thunderhill West circuit in simulation, the framework increases average overtake time from 8.8 s to 14.6 s while significantly reducing opponent progress. Meanwhile, it allows the vehicle to utilize 83.4% of available tire force. The framework achieves a 33.3 ms mean solve time (13.9 ms std), supporting real-time high-speed adversarial interaction.

Figures

Figures reproduced from arXiv: 2607.20973 by Tulga Ersal, Yijie Liao, Yufei Xi.

Figure 1
Figure 1. Figure 1: Hierarchical deep RL–MPC framework for adversarial [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Vehicle representation in the Frenet coordinate fra [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative procedurally generated training tr [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Top-down view of ego (blue) and opponent (red) trajec [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Trajectory-level comparison of ego–opponent inter [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Probability distribution of solve times. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 5 linked inside Pith

  1. [1]

    Formulation of defensive blocking in high-speed, friction-limited autonomous racing as a learned spatial- reference generation problem, addressing competitive interaction beyond lap-time optimization or overtaking

  2. [2]

    The remainder of this paper is organized as follows

    Embedding of the learned defensive intent into a non- linear Spatial Envelope MPC through policy-induced spatial regularization, enabling dynamically feasible arXiv:2607.20973v1 [cs.RO] 23 Jul 2026 and real-time closed-loop defensive control. The remainder of this paper is organized as follows. Sec. II reviews related work in autonomous racing, game-theor...

    Pith/arXiv arXiv 2026

  3. [3]

    This module is implemented using Soft Actor-Critic (SAC), where both the actor and critic are parameterized by multi-layer neural networks

    Strategic layer: A deep RL module learns high- level spatial policies in the Frenet coordinate domain, capturing the adversarial interaction and spatial structure of defensive driving. This module is implemented using Soft Actor-Critic (SAC), where both the actor and critic are parameterized by multi-layer neural networks. The actor produces a continuous ...

  4. [4]

    A spatial envelope MPC optimizes the control inputs to the vehicle to realize the learned behavior under physical feasibility and safety constraints

    Tactical layer:The learned Frenet-space policy is trans- formed into the global coordinate frame, establishing the theoretical connection between spatial references and time-parameterized trajectories. A spatial envelope MPC optimizes the control inputs to the vehicle to realize the learned behavior under physical feasibility and safety constraints

  5. [5]

    During this rollout, several metrics are evaluated to quantify task performance

    Operational layer: In this layer, the control inputs are applied to a dynamics simulator to obtain the next vehicle state. During this rollout, several metrics are evaluated to quantify task performance. Among them, collision indicates whether the ego vehicle makes con- tact with the opponent, progress measures the vehicle’s forward advancement along the ...

  6. [6]

    , Δ/u1D460/u1D441/u1D45D, Δ/u1D4511,

    State and Action Space: The state vector /u1D743/u1D461captures the essential interaction dynamics between the ego ( /u1D452) and opponent (/u1D45C) vehicles in the Frenet frame: /u1D743/u1D461= [ /u1D460/u1D452, /u1D451/u1D452, Δ/u1D713/u1D452, /u1D463/u1D452 /u1D460, /u1D463/u1D452 /u1D451, /u1D460/u1D45C, /u1D451/u1D45C, Δ/u1D713/u1D45C, /u1D463/u1D45C...

  7. [7]

    SAC Optimization: The agent is trained using standard SAC objectives. The actor network is optimized by maximizing the entropy-regularized expected return ( /u1D43D/u1D70B), while the critic network minimizes the Bellman error ( /u1D43D/u1D444) to stabilize value estimation: /u1D43D/u1D70B= E/u1D743/u1D461∼D [ /u1D6FClog /u1D70B/u1D719( /u1D741/u1D461|/u1...

  8. [8]

    Reward Function: The reward function /u1D45F/u1D461 balances progress, defensive effectiveness, and safety via fixed weig ht- ing coefficients /u1D706/u1D456> 0: /u1D45F/u1D461= /u1D7061/u1D45Fprog + /u1D7062/u1D45Fdef − /u1D7063/u1D45Fcoll − /u1D7064/u1D45Fbound (11) a) Longitudinal progress: Measured by the forward arc- length increment, accounting for clos...

    1970

  9. [9]

    Spatial Envelope MPC: High Performance Driving without a Reference,

    S. Yu et al., “Spatial Envelope MPC: High Performance Driving without a Reference,” arXiv:2509.18506, 2025

    arXiv 2025

  10. [10]

    A tube- MPC approach to autonomous multi-vehicle racing on high-sp eed ovals,

    A. Wischnewski, T. Herrmann, F. Werner, and B. Lohmann, “ A tube- MPC approach to autonomous multi-vehicle racing on high-sp eed ovals,” IEEE Trans. Intell. Veh. , vol. 8, no. 1, pp. 368–378, 2023

    2023

  11. [11]

    From the racetrack to the road: Real- time trajectory replanning for autonomous driving,

    J. K. Subosits and J. C. Gerdes, “From the racetrack to the road: Real- time trajectory replanning for autonomous driving,” IEEE Trans. Intell. Veh., vol. 4, no. 2, pp. 309–320, 2019

    2019

  12. [12]

    Nonlinear model predictive planning and control for high-speed autonomous vehicles on 3D terrai ns,

    S. Yu, C. Shen, and T. Ersal, “Nonlinear model predictive planning and control for high-speed autonomous vehicles on 3D terrai ns,” IFAC- PapersOnLine, vol. 54, no. 20, pp. 412–417, 2021

    2021

  13. [13]

    Game- theoretic planning for self-driving cars in multivehicle c ompetitive scenarios,

    M. Wang, Z. Wang, J. Talbot, J. C. Gerdes, and M. Schwager, “Game- theoretic planning for self-driving cars in multivehicle c ompetitive scenarios,” IEEE Trans. Robot. , vol. 37, no. 4, pp. 1313–1325, 2021

    2021

  14. [14]

    A game theoretic method for two-team multi-player autonomous racing,

    Z. Hu, X. Li, M. Meng, and S. Zhao, “ A game theoretic method for two-team multi-player autonomous racing,” IEEE Robot. Autom. Lett. , vol. 9, no. 9, pp. 7581–7588, 2024

    2024

  15. [15]

    Fair pla y in the fast lane: Integrating sportsmanship into autonomous raci ng systems,

    Z. Huang, C. Hao, W. Zhan, J. Ma, and M. Tomizuka, “Fair pla y in the fast lane: Integrating sportsmanship into autonomous raci ng systems,” arXiv:2503.03774, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Outracing champion Gran Turismo drivers with deep reinforcement learning,

    P . R. Wurman et al. , “Outracing champion Gran Turismo drivers with deep reinforcement learning,” Nature, vol. 602, no. 7896, pp. 223–228, 2022

    2022

  17. [17]

    A simulation benchmark for autonomous racing with large-scale human data,

    A. Remonda et al. , “ A simulation benchmark for autonomous racing with large-scale human data,” Adv. Neural Inf. Process. Syst. , vol. 37, pp. 102 078–102 100, 2024

    2024

  18. [18]

    Data-driven economic NMPC using r einforce- ment learning,

    S. Gros and M. Zanon, “Data-driven economic NMPC using r einforce- ment learning,”IEEE Trans. Autom. Control, vol. 65, no. 2, pp. 636–648, 2020

    2020

  19. [19]

    ZipMPC: Compressed context-dependent MPC cos t via imitation learning,

    R. Rickenbach, A. A. Lahoud, E. Schaffernicht, M. N. Zeil inger, and J. A. Stork, “ZipMPC: Compressed context-dependent MPC cos t via imitation learning,” arXiv:2507.13088, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Differentiable MPC for end-to-end planning and control,

    B. Amos, I. Jimenez, J. Sacks, B. Boots, and J. Z. Kolter, “Differentiable MPC for end-to-end planning and control,” Adv. Neural Inf. Process. Syst., vol. 31, 2018

    2018

  21. [21]

    Neural netw ork model predictive motion control applied to automated driving wit h unknown friction,

    N. A. Spielberg, M. Brown, and J. C. Gerdes, “Neural netw ork model predictive motion control applied to automated driving wit h unknown friction,” IEEE Trans. Control Syst. Technol. , vol. 30, no. 5, pp. 1934– 1945, 2022

    1934

  22. [22]

    Aggressive deep driving: Combining convolutional neural networks and model predictive control,

    P . Drews, G. Williams, B. Goldfain, E. A. Theodorou, and J. M. Rehg, “ Aggressive deep driving: Combining convolutional neural networks and model predictive control,” in Conference on Robot Learning , vol. 78, 2017, pp. 133–142

    2017

  23. [23]

    Online le arning of MPC for autonomous racing,

    G. Costa, J. Pinho, M. A. Botto, and P . U. Lima, “Online le arning of MPC for autonomous racing,” Robot. Auton. Syst. , vol. 167, p. 104469, 2023

    2023

  24. [24]

    Think deep and fast: Learning neural nonlinear o pinion dynamics from inverse dynamic games for split-second inter actions,

    H. Hu, J. F. Fisac, N. E. Leonard, D. Gopinath, J. DeCastr o, and G. Rosman, “Think deep and fast: Learning neural nonlinear o pinion dynamics from inverse dynamic games for split-second inter actions,” in International Conference on Robotics and Automation , 2025, pp. 16 678–16 684

    2025

  25. [25]

    Au- tonomous overtaking in Gran Turismo Sport using curriculum re- inforcement learning,

    Y . Song, H. Lin, E. Kaufmann, P . D¨ urr, and D. Scaramuzza , “ Au- tonomous overtaking in Gran Turismo Sport using curriculum re- inforcement learning,” in International Conference on Robotics and Automation, Xi’an, China, 2021, pp. 9403–9409

    2021

  26. [26]

    Blending data-driven priors in dynamic games,

    J. Lidard et al. , “Blending data-driven priors in dynamic games,” arXiv:2402.14174, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Learning to plan, planning to learn: Adaptive hierarchical RL-MPC for sample-efficient decision making,

    T. Hori, J. DeCastro, D. Gopinath, A. Balachandran, and G. Rosman, “Learning to plan, planning to learn: Adaptive hierarchical RL-MPC for sample-efficient decision making,” arXiv:2512.17091, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    A real-time terrain-adaptive local trajectory planner fo r high-speed autonomous off-road navigation on deformable terrains,

    S. Yu, C. Shen, J. Dallas, B. I. Epureanu, P . Jayakumar, a nd T. Ersal, “ A real-time terrain-adaptive local trajectory planner fo r high-speed autonomous off-road navigation on deformable terrains,” IEEE Trans. Intell. Transp. Syst. , 2024

    2024

  29. [29]

    Seitenkr ¨afte am rollenden Luftreifen,

    E. Fiala, “Seitenkr ¨afte am rollenden Luftreifen,”VDI Zeitschrift, vol. 96, pp. 973–979, 1954

    1954

  30. [30]

    Soft act or-critic: Off- policy maximum entropy deep reinforcement learning with a s tochastic actor,

    T. Haarnoja, A. Zhou, P . Abbeel, and S. Levine, “Soft act or-critic: Off- policy maximum entropy deep reinforcement learning with a s tochastic actor,” in International Conference on Machine Learning , 2018, pp. 1861–1870

    2018

  31. [31]

    Adaptable high-speed model predictive control for autonomous drifting: Koopman-based dynamics,

    M. Suminaka, J. Dallas, M. Thompson, M. Soga, E. Kasai, a nd J. Su- bosits, “ Adaptable high-speed model predictive control for autonomous drifting: Koopman-based dynamics,” in American Control Conference , 2025, pp. 1057–1063

    2025