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

The gist: A racing-to-recovery transfer learning method, which builds a recovery controller by pretraining a policy to race with drifting and then adapting it to recovery, recovers control more reliably than a policy trained on recovery alone, and more reliably than classical control baselines.

Problem Formulation

The goal is to solve the control task of returning a vehicle that has lost control while tracking a straight reference path, such as a driving lane, to a safe operating regime<ref:2610.20873#pg2> Formulating this task requires two components: a vehicle and tire model, which we describe in Section III-A, and a rigorous definition of the recovered state, described in Section III-B<ref:2610.20873#pg5> A drifting state is characterized by a large sideslip angle β, the angle between the velocity vector at the CG and the chassis centreline, and a large yaw rate Ψ˙, the rate of rotation about the vertical axis<ref:2610.20873#pg5> Recovery from an unstable vehicle state requires guiding the vehicle away from the unstable regime and towards a stable equilibrium corresponding to conventional driving<ref:2610.20873#pg5> A vehicle state is recovered if and only if zt ∈ R, given by R = δ ≤ δth, β ≤ βth, Ψ˙ ≤ Ψ˙th, β˙ ≤ β˙th, Ψ¨ ≤ Ψ¨th, eΨ ≤ eΨ,th

Controller Design and Training

The central insight is that the answer lies in racing<ref:2610.20873#pg3> A policy that has learned to race by drifting has, in the process, learned to manage saturated tires and large sideslip angles across the whole regime, which is precisely the competency that recovery requires<ref:2610.20873#pg5> We therefore investigate the following hypothesis: a policy first trained to race with drifting, and then adapted to recovery, recovers control more reliably than a policy trained on recovery alone, and more reliably than classical control baselines<ref:2610.20873#pg5> The training involves three main steps:

1. Training a racing policy with reinforcement learning that learns drifting as a time-optimal maneuver<ref:2610.20873#pg5>

2. Adapting it to recovery through continued training with a minimum-time reward<ref:2610.20873#pg5>

3. Deploying the result on a physical 1:10-scale F1TENTH vehicle using system identification, actuator-lag modelling, and domain randomization<ref:2610.20873#pg5>

Racing Pretraining and Transfer Learning

The racing task is defined as maximizing progress along a reference centreline without leaving the track, rewarding per-step progress in Frenet arc-length with a penalty for leaving the track boundaries<ref:2610.20873#pg5> The policy πrace 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<ref:2610.20873#pg5> The racing pretraining is a key component behind the superior performance of πrace→rec, as shown by ablation studies<ref:2610.20873#pg5> The transfer retains the learned control representation and discards what was fitted to the racing reward<ref:2610.20873#pg5> The actor is kept unchanged: its mapping from vehicle and track state to steering and velocity commands is the knowledge to be reused<ref:2610.20873#pg5>

Recovery Metrics and Validation

The recovery task is formalized by defining a rigorous, phase-plane-grounded definition of vehicle recovery that formalizes the “recovered” state as a tolerance region about a stable equilibrium<ref:2610.20873#pg5> The primary objective is to maximize the measure of initial conditions in U from which the vehicle can be successfully recovered, and the secondary objective is to minimize recovery time T<ref:2610.20873#pg5> Recovery metrics are defined as:

Overall recovery rate (RR)

RR = n/Ntot, where n denotes the number that successfully reach the recovered set R before truncation<ref:2610.20873#pg5>

Recovery time (RT)

RT = 1/n Xn i=1, where T is the first time at which R is reached<ref:2610.20873#pg5>

Hardware Validation and Results

Validation on a physical F1TENTH vehicle was enabled by a sim-to-real pipeline of system identification, actuator-lag modelling, and domain randomization<ref:2610.20873#pg5> The transferred policy outperforms the Stanley baseline on every tested loss-of-control state, including recovering 50% of the time from a hard state the baseline never recovers from<ref:2610.20873#pg5> In simulation, over a grid of 294 initial states more extreme than any seen in training, the transferred policy improved recoveries by 186.8% over a Stanley path tracker, exceeding both a nonlinear MPC and the recovery-only policy<ref:2610.20873#pg5> The ablation showed that racing pretraining is the essential element behind the superior recovery rate<ref:2610.20873#pg5> On the vehicle, even though the same set of pre-recorded control commands is used, the resulting initial state, z0, may slightly vary between runs<ref:2610.20873#pg5> The transfer policy outperforms the Stanley baseline on every one of three real initial states over around 180 runs, including the hardest state, from which the baseline never recovered<ref:2610.20873#pg5>

Conclusion

The paper concludes that transferring from racing is what lets a policy generalize to the broadest range of loss-of-control states, and that this advantage survives sim-to-real transfer<ref:2610.20873#pg5> The transfer policy outperformed the Stanley baseline on every one of three real initial states over around 180 runs, including the hardest state, from which the baseline never recovered<ref:2610.20873#pg5> The gap widens as the state becomes more challenging<ref:2610.20873#pg5> Overall, Fig. 4 and Table VI show that learning to race first and then fine-tuning on the recovery task outperforms every alternative that tackles the recovery problem directly<ref:2610.20873#pg5>.

--- Page 1 ---

PREPRINT SUBMITTED TO IEEE TRANSACTIONS ON INTELLIGENT VEHICLES 1

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

Teodor Ilie, Kleber Cabral, and Sidney Givigi

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<ref:2610.20873#pg5> Expert drivers regain authority in these conditions by drifting<ref:2610.20873#pg5> 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<ref:2610.20873#pg5> 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<ref:2610.20873#pg5> We define recovery relative to a phaseplane-grounded stable equilibrium and introduce Gym-Khana for controller training and evaluation<ref:2610.20873#pg5> Across 294 initial lossof-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<ref:2610.20873#pg5> An ablation shows that racing pretraining, rather than broader initial-state exposure alone, drives this gain<ref:2610.20873#pg5> 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<ref:2610.20873#pg5>Index Terms—Reinforcement learning, transfer learning, vehicle dynamics, vehicle safety<ref:2610.20873#pg5>

--- Page 2 ---

PREPRINT SUBMITTED TO IEEE TRANSACTIONS ON INTELLIGENT VEHICLES 2

I.

1. Improved AI system can recover control from a broader set of loss-of-control states by learning drifting as a time-optimal maneuver first, and then adapting it to recovery through continued training with a minimum-time reward, which improves recovery by 186.8% over a Stanley path tracker.

2. Improved AI system can generalize its control capability across the entire regime of racing and drifting by utilizing racing pretraining, as an ablation shows that this approach is drives this gain rather than broader initial-state exposure alone, allowing the policy to recover more reliably than a policy trained on recovery alone.

3. Improved AI system can be deployed in real-world scenarios with reduced performance gaps by employing a sim-to-real pipeline of system identification, actuator-lag modelling, and domain randomization, which enables the transferred policy to outperform the Stanley baseline on every tested loss-of-control state on a physical vehicle.

4. Improved AI system can be rigorously compared across different recovery methods by utilizing a phaseplane-grounded definition of vehicle recovery that formalizes the 'recovered' state as a tolerance region about a stable equilibrium, allowing for the introduction of experimental recovery metrics that make recovery controllers comparable over a common domain of loss-of-control states.

5. Improved AI system can be validated against extreme and unrecoverable states by systematically testing on a grid of 294 initial states, which includes conditions more extreme than any seen in training, enabling the comparison of recovery rate and time across both recoverable and unrecoverable subsets.