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

Teodor Ilie; Kleber Cabral; Sidney Givigi

cs.ROOct 14, 2026

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

Authors: Teodor Ilie, Kleber Cabral, Sidney Givigi

Organizations: School of Computing, Queen’s University, Kingston, ON K7L 2N8, Canada

Abstract

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.

Figures & tables

Panel of Figure 1 — Fig. 1: The single-track model. The STD states (blue) include the Cartesian position $(s_{x},s_{y})$ of the CG in the global frame; the Frenet-frame pose relative to a reference path (red) is given by the cross-track error $e_{y}$ and the heading error $e_{\Psi}$ . Variables are defined in Table I .

Quantity	Symbol	Units
Variables
Cartesian coordinates of CG	$s_{x},\ s_{y}$	m
Front axle steering angle	$\delta$	rad
Velocity at CG	$v$	m/s
Vehicle heading	$\Psi$	rad
Yaw rate	$\dot{\Psi}$	rad/s

TABLE I: Single-Track Model Quantities

Fig. 2: Sideslip vs. yaw-rate phase plane of Hindiyeh’s two-state single-track model [ 2 ] .

Panel of Figure 4 — Fig. 3: Overview of the approach. 1. A policy is trained with proximal policy optimization (PPO) to race on the Drift track and its mirror image, where drifting emerges as the time-optimal technique. 2. Its actor $\pi_{\text{race}}$ is transferred unchanged to the recovery task, and its critic is reinitialized. 3. Training continues under the recovery reward, which drives the vehicle from an unrecovered state in $\mathcal{U}$ into the recovered set $\mathcal{R}$ ; the two tasks share one vehicle model, observation and action space, so only the reward changes between them. 4. The resulting $\pi_{\text{race}\to\text{rec}}$ is deployed on the physical F1TENTH vehicle. Before deployment, system identification, actuator-lag modelling and domain randomization adapt the simulator to the vehicle, and the controller is retrained on it from racing onwards (Sections IV-D and VI ). Every image is taken from our experiments: the racing and recovery panels are rollouts of the trained policies in simulation, the bottom-left plot is the loss ( 6 ) of the identification run across four parallel covariance matrix adaptation evolution strategy (CMA-ES) workers, and the bottom-right photo overlays frames of a recovery run on the physical vehicle. In both recovery panels time runs downwards: the shading marks $\mathcal{U}$ (red) and its complement $\mathcal{R}$ (green), the dotted line the initial state $\mathbf{z}_{0}$ where the policy takes over, and the solid line the first recovered state $\mathbf{z}_{T}$ .

Hyperparameter	Value
Parallel environments	64
Total training timesteps	$120\times 10^{6}$
Rollout length	4096 timesteps
Timestep	$0.01$ s ( $100$ Hz)
Learning rate (linear)	$3\times 10^{-4}\rightarrow 1\times 10^{-4}$
Actor network	MLP, two hidden layers of 128, tanh

TABLE II: PPO Hyperparameters Modified from SB3 Defaults

Quantity	Symbol	Range	Equivalent
Velocity	$v_{0}$	$[2,\ 12]$ m/s	—
Sideslip angle	$\beta_{0}$	$[-0.349,\ 0.349]$ rad	$\pm 20^{\circ}$
Yaw rate	$\dot{\Psi}_{0}$	$[-0.785,\ 0.785]$ rad/s	$\pm 45^{\circ}$ /s
Heading error	$e_{\Psi,0}$	$[-0.785,\ 0.785]$ rad	$\pm 45^{\circ}$

TABLE III: Initial-State Sampling Ranges for Training

Quantity	Symbol	Value	Equivalent
Steering angle	$\delta_{\mathrm{th}}$	$0.175$ rad	${\approx}\,10^{\circ}$
Sideslip angle	$\beta_{\mathrm{th}}$	$0.07$ rad	${\approx}\,4^{\circ}$
Yaw rate	$\dot{\Psi}_{\mathrm{th}}$	$0.175$ rad/s	${\approx}\,10^{\circ}$ /s
Sideslip rate	$\dot{\beta}_{\mathrm{th}}$	$7.0$ rad/s	—
Yaw acceleration	$\ddot{\Psi}_{\mathrm{th}}$	$17.5$ rad/s 2	—
Heading error	$e_{\Psi,\mathrm{th}}$	$0.087$ rad	${\approx}\,5^{\circ}$

TABLE IV: Recovered-Set Thresholds

Variable	Symbol	Range	Points
Sideslip angle (rad)	$\beta_{0}$	$\pm 1.39$	7
Yaw rate (rad/s)	$\dot{\Psi}_{0}$	$\pm 13$	7
Velocity (m/s)	$v_{0}$	$[4,\ 5]$	2
Heading error (rad)	$e_{\Psi,0}$	$\pm 0.17$	3

TABLE V: Evaluation Grid Over the Four Initial-State Variables

Fig. 4: Recovery rate $\mathrm{RR}$ across the initial states of the nominal evaluation grid of Table V . Each cell is the percentage of its six $(v_{0},e_{\Psi,0})$ episodes that recover, so it takes one of seven values; cells with no recoveries are left unlabelled, and the dashed lines mark $\beta_{0}=0$ and $\dot{\Psi}_{0}=0$ . Each panel is one row of Table VI , named and marked (a)–(d) as there, and its $\mathrm{RR}$ is that row’s. The Stanley baseline performs poorest, the policies without racing pretraining and without recovery fine-tuning similarly well, and the transfer policy best.

		Overall				Stanley-rec. ( $\mathcal{SR}$ )		Stanley-unrec. ( $\mathcal{SU}$ )
Controller	Fig. 4	$\mathrm{RI}\,\uparrow$	$\mathrm{RR}\,\uparrow$	$n\,\uparrow$	$\mathrm{RT}\,\downarrow$	$n_{\text{SR}}\,\uparrow$	$\mathrm{RT}_{\text{SR}}\,\downarrow$	$n_{\text{SU}}\,\uparrow$	$\mathrm{RT}_{\text{SU}}\,\downarrow$
		(%)	(%)	(of 294)	(s)	(of 38)	(s)	(of 256)	(s)
Baseline controllers
Stanley	(a)	—	12.9	38	0.789	—	0.789	—	—
ST-MPC		$+13.2$	14.6	43	0.813	34	0.774	9	0.961
Proposed controller

TABLE VI: Recovery Performance and Ablation on the Nominal Evaluation Grid

Fig. 5: The modified F1TENTH hardware platform from varying angles, showing custom tire covers, lower ride height, harder shock absorbers, and stock F1TENTH hardware components.

Symbol	Easy	Medium	Hard
$\beta_{0}$ ( ∘ )	$-6$ ( $-9$ , $-3$ )	$-19$ ( $-24$ , $-13$ )	$28$ ( $25$ , $34$ )
$\dot{\Psi}_{0}$ ( ∘ /s)	$45$ ( $14$ , $100$ )	$92$ ( $39$ , $154$ )	$-120$ ( $-176$ , $-70$ )
$v_{0}$ (m/s)	$3.9$ ( $3.8$ , $4.1$ )	$3.5$ ( $3.3$ , $3.6$ )	$3.2$ ( $3.1$ , $3.4$ )
$e_{\Psi,0}$ ( ∘ )	$-16$ ( $-19$ , $-14$ )	$0$ ( $-2$ , $2$ )	$-36$ ( $-37$ , $-35$ )
$\delta_{0}$ ( ∘ )	$15$ ( $15$ , $15$ )	$0$ ( $-12$ , $8$ )	$0$ ( $0$ , $0$ )

TABLE VII: The Three Real-Vehicle Initial States

Easy state
Controller	$N$	$n\,\uparrow$	$\mathrm{RR}\,\uparrow$	$\mathrm{RT}\,\downarrow$	$\mathrm{RT}$ median (IQR) $\,\downarrow$
			(%)	(s)	(s)
Stanley	30	21	70.0	0.815	0.900 (0.730, 0.939)
$\pi_{\text{race}\to\text{rec}}$	30	$\mathbf{30}$	$\mathbf{100.0}$	$\mathbf{0.665}$	$\mathbf{0.605}$ (0.560, 0.707)
Medium state
Stanley	30	4	13.3	$\mathbf{0.610}$	$\mathbf{0.455}$ (0.178, 0.888)

TABLE VIII: Real-Vehicle Recovery Performance per Initial State

Panel of Figure 14 — Fig. 6: Real-vehicle recovery results for the Stanley baseline and the transfer policy $\pi_{\text{race}\to\text{rec}}$ , a qualitative view of the 179 runs of Table VIII . (a) Recovery rate. (b) Recovered runs are plotted at their recovery time, with the mean $\mathrm{RT}$ marked by a black bar; failed runs are shown as crosses in the band below. Box plots show $\mathrm{RT}$ distributions, and are omitted on the Stanley medium and hard states where there are too few successful runs for a meaningful measurement.

RL for Autonomous Driving

Reinforcement learning for autonomous driving trains vehicle policies through interaction to plan, control, and coordinate safely in traffic.

See all 120 papers

Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints
Mar 6, 2026Bo Leng, Weiqi Zhang, Zhuoren Li +4Constrained RL Autonomous Racing
College of Automotive and Energy Engineering, Tongji University, Shanghai, 201804, China · School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, 639798, Singapore
Learning from Mistakes: Rollout-Retrieval Lifelong Policy Learning for Autonomous Driving
Jun 29, 2026Cheng Gong, Haoyang Wang, Chao Lu +2Autonomous Driving Planning Continual RL
School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081, China · School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798
CRRL: A Causality-Based Reinforcement Learning Framework for Autonomous System Recovery
Jul 3, 2026Safia Fatima, Kai Olav Ellefsen, Leon MoonenReinforcement Learning Robot Failure Recovery
Simula Research Laboratory, Norway · University of Oslo, Norway

Mar 6, 2026J/K move · Enter open · S save

Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints

Bo Leng, Weiqi Zhang, Zhuoren Li, Lu Xiong+3

College of Automotive and Energy Engineering, Tongji University, Shanghai, 201804, China · School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, 639798, Singapore

Constrained RL Autonomous Racing

Open paper AskPDF arXiv

Reinforcement learning has shown significant potential for autonomous racing, but it still faces challenges such as training instability, inefficient exploration, and unsafe action outputs in high-dynamic racing scenarios. This paper proposes a Trajectory guidance and Dynamics constraints Reinforcement Learning (TraD-RL) framework for autonomous racing. The proposed method incorporates expert prior knowledge into policy learning through Minimum Curvature Racing Line (MCRL) guidance, explicit vehicle dynamics constraints, and two-stage curriculum learning. MCRL provides global path and velocity references through observation augmentation and reward shaping, thereby improving exploration efficiency and racing performance. Yaw rate and sideslip angle constraints are introduced to characterize the vehicle dynamic safe operating envelope, and the corresponding stability costs are incorporated into policy optimization through Lagrangian relaxation. Moreover, the two-stage curriculum learning strategy enables a progressive transition from stable trajectory following to high-speed performance exploration. Experiments on two racetracks demonstrate that TraD-RL improves racing performance while maintaining a favorable balance between speed and dynamic stability. Further analyses of ablation, sensitivity, and robustness validate the effectiveness and stability of the proposed framework.

Abstract

RL for Autonomous Driving

Explore similar work

Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints

Learning from Mistakes: Rollout-Retrieval Lifelong Policy Learning for Autonomous Driving

CRRL: A Causality-Based Reinforcement Learning Framework for Autonomous System Recovery