Safety with Agency: Human-Centered Safety Filter with Application to AI-Assisted Motorsports

TL;DR

This work tackles safety in human–AI shared autonomy for high-speed motorsports by introducing a fully model-free human-centered safety filter (HCSF). The approach learns a neural state–action safety value function and a best-effort fallback policy via model-free RL and enforces a novel state–action safety constraint ${\boldsymbol{\mathcal{Q}}}(x,u)\geq\gamma V(x)$ without requiring system dynamics. In a large, in-person study with a high-fidelity racing simulator, HCSF improves safety and user satisfaction while preserving human agency and comfort, outperforming conventional safety filters and unassisted driving. The results demonstrate scalable, transparent human–robot collaboration in complex, black-box environments and suggest broader applicability to safety-critical shared autonomy settings beyond racing.

Abstract

We propose a human-centered safety filter (HCSF) for shared autonomy that significantly enhances system safety without compromising human agency. Our HCSF is built on a neural safety value function, which we first learn scalably through black-box interactions and then use at deployment to enforce a novel state-action control barrier function (Q-CBF) safety constraint. Since this Q-CBF safety filter does not require any knowledge of the system dynamics for both synthesis and runtime safety monitoring and intervention, our method applies readily to complex, black-box shared autonomy systems. Notably, our HCSF's CBF-based interventions modify the human's actions minimally and smoothly, avoiding the abrupt, last-moment corrections delivered by many conventional safety filters. We validate our approach in a comprehensive in-person user study using Assetto Corsa-a high-fidelity car racing simulator with black-box dynamics-to assess robustness in "driving on the edge" scenarios. We compare both trajectory data and drivers' perceptions of our HCSF assistance against unassisted driving and a conventional safety filter. Experimental results show that 1) compared to having no assistance, our HCSF improves both safety and user satisfaction without compromising human agency or comfort, and 2) relative to a conventional safety filter, our proposed HCSF boosts human agency, comfort, and satisfaction while maintaining robustness.

Figures (16)

• Figure 1: Our proposed enables robust and smooth safety interventions for shared autonomy systems. (a) switches to the best-effort fallback policy at the last possible moment. However, this switching can feel abrupt and uncomfortable for human operators. (b) Our instead intervenes smoothly while promoting human agency, thereby reducing automation surprise and enhancing user experience. (c) Users interact with a high-fidelity racing simulator via a steering wheel and set of pedals (throttle and brake).
• Figure 2: Illustration of our intervention at a hairpin corner (i.e., a sharp turn requiring rapid deceleration). Without safety filter assistance, inexperienced human drivers often miss the braking point, leading to understeering and the vehicle leaving the track. In contrast, our monitors the state and the human action to determine the braking point and provides necessary steering and braking interventions that keep the vehicle on the track. Braking assistance is visible through the rear lights.
• Figure 3: A diagram describing the interaction between a human operator, our proposed , and game environment. Our utilizes a safety value function that we learn scalably through black-box interactions via model-free -based reachability analysis, and at runtime leverages our novel constraint to enforce safety without any knowledge of system dynamics. Moreover, it intervenes minimally and smoothly to enhance human agency and comfort. Finally, our communicates the action modifications to the human driver via visual cues, facilitating transparent human--robot collaborartion.
• Figure 4: Our HCSF displays two types of visual cues: horizontal arrows that reflect the modifications made to the steering input, and vertical arrows that indicate the corrections made to the throttle/braking inputs. The length of each arrow is proportional to the magnitude of modification made to the corresponding input channel.
• Figure 5: Our achieved near-zero failures throughout the user study, demonstrating significant enhancement in safety compared to unassisted human driving. Although our outperformed in both failure modes, the differences were not statistically significant. Statistical significance is marked with asterisks, where more asterisks indicate larger significance.

Theorems & Definitions (15)

• Definition 1: Discrete-time CBF agrawal2017discrete
• Remark 1
• Proposition 1
• proof
• Definition 2: Human-Centered Safety Filter
• Proposition 2: Recursive Feasibility of HCSF
• proof
• Lemma 1
• proof
• Lemma 2