Preprint ·
TriDrive: a real-time cockpit world model, now on arXiv
TriDrive jointly forecasts driver kinematics, vehicle dynamics and road demands from one automation-conditioned transition model, runs at 5 Hz on a comma four, and was validated in a 14-participant on-road study.
TriDrive: Joint Driver, Vehicle, and Road Modeling for Forecasting and Driver Monitoring (Yuhang Wang, Jingxin Yang, Chuheng Wei, Yuechen Guo, Jinghan Xu, Zhao Han, Hao Zhou) is on arXiv. One transition model forecasts the driver, the vehicle and the road together; it reaches 48.05 All-MPJPE on the AIDE driver-kinematics benchmark (versus 71.47 for the previous best), runs with 177 ms p95 latency on a comma four with an external GPU, and its warnings were rated more appropriate and more timely by 14 drivers in an on-road study. Explore the three-stream forecast on the project page.