PaperlayerView exact revision on arXiv

Source-linked AI summary

Race to Recovery: Designing Advanced Safety Systems for Autonomous Driving via Learning Autonomous Racing

Teodor Ilie, Kleber Cabral, Sidney Givigi

arXiv:2610.20873v1cs.RO

TL;DR

Production safety systems are designed to prevent instability, while existing autonomous recovery work leaves gaps in scenario breadth, recovery definitions, and online operation. The paper trains a racing policy that learns drifting, adapts it to recovery, and evaluates it using a phase-plane-grounded definition. In simulation the transferred policy improves recoveries by 186.8% over Stanley, and on hardware it outperforms Stanley on each tested state.

  • Problem

    Production safety systems focus on preventing instability, and existing recovery approaches use narrow scenarios, inconsistent recovery definitions, or offline-planned trajectories.

    Source: §I · p. 1

  • Method

    The paper pretrains a racing policy to learn drifting, adapts it to recovery with continued training, and grounds recovery in a phase-plane stable equilibrium.

    Source: §I · p. 2, §VII · p. 11

  • Results

    186.8% improvement over a Stanley path tracker across 294 initial states; on hardware, the transferred policy outperformed Stanley on every tested state.

    Source: §VII · pp. 11–12

  • Takeaways & Limitations

    The results support racing pretraining as a way for a policy to generalize across a broad range of loss-of-control states, with the advantage surviving sim-to-real transfer.

    Source: §VII · p. 12

Abstract

from arXiv · show

Vehicles that lose control at the handling limit through icing, hydroplaning, or an abrupt evasive maneuver fall outside the operating envelope of production safety systems such as electronic stability control, which focus on preventing instability rather than recovering from it. Expert drivers regain authority in these conditions by drifting. Existing autonomous recovery approaches are validated on idealized scenarios, define "recovered" states empirically and inconsistently, or rely on offline-planned trajectories rather than acting online. We address these gaps by first training a racing policy with reinforcement learning that learns drifting as a time-optimal maneuver, then adapting it to recovery through continued training with a minimum-time reward. We define recovery relative to a phase-plane-grounded stable equilibrium and introduce Gym-Khana for controller training and evaluation. Across 294 initial loss-of-control states, the transferred policy improves recovery by 186.8% over a Stanley path tracker and outperforms both a nonlinear model predictive controller and a recovery-only policy trained from scratch, while also recovering faster on states the baseline can already handle. An ablation shows that racing pretraining, rather than broader initial-state exposure alone, drives this gain. Deployed on a physical 1:10-scale F1TENTH vehicle using system identification, actuator-lag modelling, and domain randomization, the transferred policy outperforms Stanley on every tested loss-of-control state, including recovering 50% of the time from a hard state the baseline never recovers from.

I. INTRODUCTION

The paper targets recovery after a vehicle leaves the stable handling envelope, where production systems designed to prevent instability may not help. It proposes learning recovery through drifting-capable racing pretraining and evaluates a common, phase-plane-grounded definition of recovery.

I. INTRODUCTION

  • Existing recovery approaches address narrow idealized scenarios, define recovery inconsistently, or rely on offline-planned trajectories rather than acting online.The paper frames these as gaps in establishing recoverable-state breadth, comparable outcomes, and online intervention.Source: §I · p. 1
  • The method pretrains a racing policy that learns to drift, then adapts it to recovery through continued training with a minimum-time reward.The paper argues racing teaches management of saturated tires and large sideslip angles, which recovery also requires.Source: §I · pp. 1–2
  • Across 294 initial loss-of-control states, the transferred policy improved recoveries by 186.8% over a Stanley path tracker and outperformed MPC and a recovery-only policy trained from scratch.The states spanned conditions more extreme than any seen in training.Source: §I · p. 2
  • The paper defines recovery around a phase-plane-grounded stable equilibrium and introduces Gym-Khana for training and evaluating controllers at the handling limit.The definition formalizes recovery as a tolerance region about a stable equilibrium and supports comparison over a common domain of loss-of-control states.Source: §I · p. 2
  • On a physical 1:10-scale F1TENTH vehicle, the transferred policy outperformed Stanley across 179 runs from three real loss-of-control initial states.The sim-to-real pipeline uses system identification, actuator-lag modelling, and domain randomization.Source: §I · p. 2

II. RELATED WORK

Prior work moves from model-based optimal control toward learning for control at the handling limit, while recovery research has begun using drifting but leaves gaps in scope, comparability, and online operation.

A. Control at the Limit of Handling: From MPC to RL

  • Combined-slip MPC formulations face nonconvex optimization at each control step, limiting model fidelity, horizon length, and control rate, while requiring hand-tuned costs.Reinforcement learning has demonstrated racing and drifting results beyond optimal-control approaches, including performance over expert humans and state-of-the-art MPC.Source: A. Control at the Limit of Handling:… · p. 2
  • Learning-based racing and drifting controllers have demonstrated strong performance, and drifting can emerge as the fastest way around a track from progress and out-of-bounds rewards.Prior sim-to-real drifting work combined actuator-lag modelling, learning-based system identification, and targeted domain randomization.Source: A. Control at the Limit of Handling:… · p. 2

B. Drifting for Active Safety and Recovery

  • Prior recovery studies cover specific scenarios, use differing empirical recovery criteria, or depend on offline planning for real-vehicle validation.Examples include collision stabilization, a planned drift-based U-turn, drifting MPC for collision avoidance, and learned recovery from oversteer.Source: B. Drifting for Active Safety and… · pp. 2–3

III. PROBLEM FORMULATION

The formulation seeks to return an uncontrolled vehicle tracking a straight reference path to a safe operating regime. Its model must capture beyond-linear tire dynamics while remaining interpretable and identifiable from physical-vehicle measurements.

III. PROBLEM FORMULATION

  • The recovery task is returning a vehicle that has lost control while tracking a straight reference path to a safe operating regime.The formulation requires a vehicle-and-tire model and a rigorous definition of the recovered state.Source: §III · p. 3

A. Vehicle and Tire Model

  • The vehicle model must represent tire dynamics beyond the linear handling regime while remaining interpretable and identifiable from physical-vehicle measurements.The authors identify the physical vehicle in Section IV-D.Source: A. Vehicle and Tire Model · p. 3
  • Tire lateral force rises with slip angle, peaks near the friction limit, then decays in saturation; combined-slip forces share a friction-circle bound.Longitudinal force spent at a tire reduces the lateral force it can still generate, and vice versa.Source: A. Vehicle and Tire Model · p. 3
  • The single-track model represents global Cartesian position and Frenet-frame pose relative to the reference path, including cross-track and heading errors.The figure distinguishes Cartesian states in blue from Frenet-frame quantities in red.Source: Figure 1 · p. 3, A. Vehicle and Tire Model · pp. 3–4
  • Drifting combines large sideslip and yaw rate with saturated driven tires and counter-steer, whose throttle and steering inputs act on yaw moment in opposite directions.Throttle increases longitudinal force at the driven tires, while counter-steer reduces front lateral force and yaw moment.Source: A. Vehicle and Tire Model · p. 3
  • The authors use a single-track chassis with combined-slip Pacejka tires; front and rear wheel angular velocities let it resolve longitudinal slip and combined tire forces.The model has a nine-element state vector and is driven by steering rate and longitudinal acceleration.Source: A. Vehicle and Tire Model · p. 3

B. The Recovery Task

Recovery is grounded in the vehicle’s stable straight-driving equilibrium, with practical bounds defining when the vehicle counts as recovered. The task is to reach that set without exceeding a lateral tracking limit, prioritizing the breadth of recoverable initial states and then recovery time.

B. The Recovery Task

  • The recovered set is grounded in the stable zero-sideslip, zero-yaw-rate equilibrium for conventional straight driving.The phase plane distinguishes this stable equilibrium from drift equilibria, which are saddle points requiring active steering and throttle to hold.Source: B. The Recovery Task · p. 4
  • Recovery requires dynamic-state bounds and path-alignment bounds, with derivative limits ensuring the stable state is held rather than crossed transiently.The first five conditions constrain the dynamic state; the final two constrain alignment and proximity to the reference path.Source: B. The Recovery Task · p. 4
  • The non-recovered set is unbounded and partly unrecoverable, and the boundary of its recoverable portion is unknown in closed form.Source: B. The Recovery Task · p. 4
  • Recovery succeeds when the vehicle reaches the recovered set without violating the lateral tracking constraint during the maneuver.The task prioritizes maximizing the measure of successfully recoverable initial states, then minimizing recovery time.Source: B. The Recovery Task · p. 4
  • Fig. 3 depicts recovery rollouts with time running downward, shading unrecovered states U red and recovered states R green, and marking the initial and first recovered states.Source: Figure 3 · p. 5

IV. RECOVERY CONTROLLER DESIGN

The controller design uses a shared vehicle model and policy interface to train for racing and recovery, then transfers a racing policy that discovers drifting through trial and error. Its observations combine vehicle state with local track geometry, while its actions command steering and velocity.

IV. RECOVERY CONTROLLER DESIGN

  • The design trains a racing policy in Gym-Khana, transfers it to recovery, and validates the controller on a physical F1TENTH vehicle.Racing pretraining explores vehicle behaviors including initiating, sustaining, and exiting a drift.Source: §IV · p. 5

A. Observations and Actions

  • The policy observes vehicle state and upcoming track geometry, and outputs commanded steering angle and velocity.The track lookahead uses five uniformly spaced centreline points at 0.5 m spacing, with curvature at each point and width at the first and last points.Source: A. Observations and Actions · pp. 5–6

B. Racing Pretraining

  • Racing maximizes centreline progress without leaving the track, allowing the policy to discover drifting as a time-optimal cornering technique on a low-friction surface.No raceline or driving style is prescribed; per-step progress serves as a surrogate for lap time.Source: B. Racing Pretraining · p. 6
  • Racing episodes terminate when the vehicle leaves the track, truncate after 4096 timesteps, and reset it at rest at a uniformly sampled centreline position.The timestep limit bounds training rollouts rather than defining the task.Source: B. Racing Pretraining · p. 6

C. Racing-to-Recovery Transfer

The recovery controller reuses the racing policy’s actor and continues training under a recovery reward, while resetting components fitted to the racing objective. Recovery episodes sample perturbed initial states and aim both to recover more states and to do so in less time.

C. Racing-to-Recovery Transfer

  • The racing actor is transferred unchanged to recovery, while the critic is reinitialized and training continues under the recovery reward.The transfer retains the learned control representation but discards what was fitted to the racing reward.Source: C. Racing-to-Recovery Transfer · p. 6
  • Recovery episodes sample initial speed, sideslip, yaw rate, and heading error uniformly from specified ranges, and end on recovery-set entry or lane departure.Episodes also truncate after 2048 timesteps or 46 m of path travel, with bootstrapping at truncation but not at terminal states.Source: C. Racing-to-Recovery Transfer · p. 6
  • The recovery reward targets both recovering from more initial states and minimizing recovery time, with a per-step penalty serving the time objective.The out-of-bounds penalty and recovery reward serve the objective of recovering successfully.Source: C. Racing-to-Recovery Transfer · p. 6
  • The transfer resets the critic, optimizer moments, learning-rate schedule, and action-noise scale before 50 million further training timesteps.Action noise is reset to log σ = −0.5 from the converged racing value; the racing run began at −1.0.Source: C. Racing-to-Recovery Transfer · p. 6

D. Sim-to-Real Transfer

Sim-to-real adaptation fits unmeasured simulator parameters to vehicle data, models actuator lag, and applies targeted domain randomization. The resulting identified parameters are checked against held-out trajectories, while randomized parameters are used in training but not evaluation.

D. Sim-to-Real Transfer

  • System identification fits 29 unmeasured chassis, actuator, and tire parameters by replaying manual vehicle inputs in simulation.Directly measurable quantities, including mass, dimensions, torque split, and steering limits, are measured separately.Source: D. Sim-to-Real Transfer · p. 7
  • The fit minimizes weighted mean-squared error between simulated and real signals over short open-loop windows, then validates parameters on held-out trajectories.The simulator is treated as a black box and optimized with CMA-ES; the channels include velocities, yaw rate, wheel speed, pose, and sideslip.Source: D. Sim-to-Real Transfer · p. 7
  • Steering and velocity actions are modeled with first-order actuator lag, with time constants identified as part of the parameter vector.Source: D. Sim-to-Real Transfer · p. 7
  • Domain randomization targets uncertain actuator parameters and four influential tire parameters, using ±5% perturbations clipped at ±15%.Randomized values are held fixed within each training episode; evaluation uses nominal parameters for comparable metrics.Source: D. Sim-to-Real Transfer · p. 7

V. SIMULATION EXPERIMENTS

The simulation protocol compares learned recovery controllers with Stanley and single-track MPC on a broad grid of initial loss-of-control states. It measures recovery efficacy and time relative to baseline performance.

A. Baselines and Metrics

  • Stanley tracks a straight path using heading and cross-track errors, but its steering-only control holds speed at the initial velocity.The single-track MPC penalizes sideslip but cannot use the learned policies’ full combined-slip tire model because of model complexity.Source: A. Baselines and Metrics · p. 7
  • The authors limit their comparison claim to conservative-handling baselines, not the best drift-capable MPC.They note that RL versus state-of-the-art MPC at the handling limit has already been addressed by prior work.Source: A. Baselines and Metrics · p. 7
  • The 294-state evaluation grid covers combinations of four initial-state variables and extends beyond the learned policies’ training distribution.Its sideslip and yaw-rate bounds were expanded until the most extreme grid corners were unrecoverable for all controllers.Source: A. Baselines and Metrics · p. 8, Table V · p. 8
  • Recovery improvement measures the percentage increase in total recoveries over Stanley, while recovery times and counts are also compared on its 38 recoverable and 256 unrecoverable states.Recovery times are compared fairly only on states both controllers recover, because successful-only averages can include harder states for the more effective controller.Source: A. Baselines and Metrics · p. 8

B. Simulation Results

The transferred racing-to-recovery policy leads the simulation comparisons in recovery efficacy and is faster than Stanley on the states Stanley can recover. Stanley’s successes cluster where yaw rate is already moving the vehicle toward recovery.

B. Simulation Results

  • 186.8% recovery improvement over Stanley; the transfer policy recovers 37 of 38 Stanley-recoverable states and 72 of 256 Stanley-unrecoverable states.On the shared Stanley-recoverable set, its mean recovery time is 0.347 s versus Stanley’s 0.789 s.Source: B. Simulation Results · p. 8
  • Stanley succeeds mainly where yaw rate is already rotating the vehicle out of sideslip and toward recovery unaided.Source: B. Simulation Results · p. 9

C. Ablation

The ablation identifies racing pretraining as the key contributor to recovery rate, while recovery-only training can be faster. Hardware validation uses a modified F1TENTH vehicle and a retrained policy with identified vehicle dynamics.

C. Ablation

  • The transfer policy recovered from 109 states, versus 95 for racing-only and 94 for recovery-only training; the comparison favors racing pretraining over direct recovery alternatives.The authors conclude that racing pretraining is essential to the superior recovery rate, while wider initial-state exposure alone does not close the gap.Source: C. Ablation · p. 9, Table VI · p. 10, Figure 4 · p. 10
  • Recovery-only policies trained on wider initial-state ranges reached RI = 128.9% and 142.1%, neither improving on πrec or matching the transfer policy.The wider-range and curriculum variants indicate that broader exposure alone is not a substitute for racing pretraining.Source: C. Ablation · p. 9
  • πrec had RT = 0.450 s, faster than the transfer policy at 0.740 s and πrace at 1.227 s, reflecting a recovery-rate versus recovery-time trade-off.The authors attribute the likely difference to recovery-only training specializing in straight-line minimum-time recovery, while racing-style drifting may improve efficacy in harder scenarios.Source: C. Ablation · p. 9

A. Testbed and Deployment

  • The physical testbed is a modified 1:10-scale F1TENTH car with harder shocks, lower ride height, faster steering, and tire covers tested on a smooth epoxy floor.The tire–road combination was selected to retain controllability while making recovery challenging at experimental speeds.Source: A. Testbed and Deployment · p. 9

VI. HARDWARE VALIDATION

  • The racing-to-recovery policy was retrained using identified vehicle parameters and modeled actuator dynamics in Gym-Khana before physical evaluation.The hardware-validation setup replaces nominal simulation parameters with parameters identified on the physical vehicle.Source: §VI · p. 9, A. Testbed and Deployment · p. 10

A. Testbed and Deployment

  • Hardware evaluation compared Stanley with the transferred policy on three selected initial states, repeating each experiment 30 times to characterize variation in physical starting states.The full 294-state grid was impractical to reproduce physically, and the initial states were selected to include mixed recoveries and failures.Source: A. Testbed and Deployment · pp. 10–11

B. Real-Vehicle Experiments

On the physical vehicle, the transferred policy recovered more often than Stanley on all three tested loss-of-control states. Its advantage was largest on the hard state, where Stanley never recovered.

B. Real-Vehicle Experiments

  • Both controllers were run 30 times on each of three initial states (easy, medium, hard), totaling 180 runs; one πrace→rec medium-state run was corrupted, leaving 179 valid runs.Source: B. Real-Vehicle Experiments · p. 11
  • The transferred policy recovered on 100% of easy-state runs versus Stanley’s 70%, and on 50% of hard-state runs versus Stanley’s 0%.Stanley never recovered from the hardest tested state; the transferred policy’s recovery advantage held across all three states.Source: B. Real-Vehicle Experiments · p. 11, Table VIII · p. 11
  • Stanley’s lower medium-state recovery time is not directly comparable because it recovered only 13.3% of runs, so the authors compare times on the common recovered-state set.Recovery-time comparisons based only on successful runs can reflect which runs each controller recovered.Source: B. Real-Vehicle Experiments · p. 11

VII. CONCLUSION

The paper transfers a drifting racing policy to recovery, grounding recovery in the β– ˙Ψ phase plane and evaluating controllers across broad loss-of-control states. The transferred policy outperforms the baselines in simulation and on the tested real-vehicle states, while validation scope and speed-versus-efficacy trade-offs remain open.

VII. CONCLUSION

  • The method trains a policy to race with drifting, then adapts it to recovery with a minimum-time reward, grounding recovered states in the β– ˙Ψ phase plane.Both tasks share a vehicle model, observation space, and action space; the controller was deployed through system identification, actuator-lag modelling, and domain randomization.Source: §VII · p. 11
  • 186.8% more recoveries than Stanley were achieved by the transferred policy across 294 simulated initial states, outperforming nonlinear MPC and a recovery-only policy.The initial states were more extreme than any seen in training, and the ablation attributed the advantage to racing pretraining rather than broader initial-state exposure alone.Source: §VII · p. 11
  • On all three tested real initial states, the transferred policy outperformed Stanley, including the hardest state, from which Stanley never recovered.Fig. 6 summarizes recovery rate and recovery-time results across 179 runs.Source: Figure 6 · p. 12, §VII · p. 12
  • Validation remains limited to a straight reference path and external motion-capture state estimation, leaving curved paths and onboard noisy estimation for future study.The paper also identifies statistical characterization of the exploited drifting maneuvers as future work.Source: §VII · p. 12
  • Real-vehicle validation on a broader set of initial states and the efficacy-versus-speed trade-off under different mission priorities remain open.The paper presents these as directions for improving applicability to real-world autonomous driving systems.Source: §VII · p. 12
Loading 2610.20873v1…