Assistive agents in shared driving learn to guess intent first, improving torque alignment and collision rate in racing and urban driving

An assistive agent in shared driving can now infer the human's objective before deciding how much to intervene, rather than only reacting to the current vehicle state. Prior assistive policies lacked explicit modeling of human intent, making it hard to be proactive and legible at the same time. Three human-subjects studies and a preliminary on-vehicle deployment cover high-speed racing (N=30) and urban driving (N=10), two qualitatively distinct domains, with intent conditioning yielding measurable improvements in torque alignment and collision rate.

The method extends a recurrent state-space world model with intent inference, so the assistive policy learns by repeatedly coordinating with a virtual partner in imagination. That training scheme is called fictitious co-play, in which the policy anticipates and supports discrete human objectives without overriding human agency. Validation settings are high-speed racing and urban driving with richer intent structure, with N=30 and N=10 respectively.

Bridging a world-model policy to a road-ready system requires closing the sim-to-real-human gap through policy down-selection and reward restructuring. This step sits inside an iterative human-machine interface design rather than a single offline training pass. The deployment pipeline is exercised in the preliminary on-vehicle deployment across the two domains.

The demonstrated scope is human-subjects studies in high-speed racing (N=30) and urban driving (N=10) plus a preliminary on-vehicle deployment, aimed at assistive agent design in shared vehicle control.

The summary states only that intent conditioning yields measurable improvements in torque alignment and collision rate without specific figures, and the on-vehicle deployment is described as preliminary.