Driver-WM: Latent World Model for In-Cabin Dynamics Rollout
The arXiv preprint 2605.05092 presents Driver-WM, a latent world model focused on drivers, which simulates in-cabin dynamics based on external traffic conditions. This model integrates the forecasting of physical kinematics with the recognition of behavioral and emotional semantics. It functions within a compact latent space derived from fixed vision-language features, employing a dual-stream architecture to process both external traffic and internal driver states, linked through a gated causal injection mechanism. This research tackles the existing gap in multi-step rollout capabilities for driver dynamics in current in-cabin intelligence systems, with the goal of enhancing safety in L2/L3 driving automation during shared-control transitions.
Key facts
- arXiv:2605.05092
- Driver-WM is a driver-centric latent world model
- Rolls out in-cabin dynamics conditioned on out-cabin traffic context
- Unifies physical kinematics forecasting with behavioral and emotional semantic recognition
- Operates in compact latent space from frozen vision-language features
- Dual-stream architecture encodes external traffic and internal driver states
- Gated causal injection mechanism couples the two streams
- Addresses lack of multi-step rollout for driver dynamics in in-cabin intelligence
Entities
Institutions
- arXiv