Event
OptCar: Specializing a General FKD Model with Small Amounts of Real-Vehicle Data
1 reports · 1 sources
Rwik Rana et al. propose the OptCar method, which specializes a general forward kinematics-dynamics (FKD) model into a high-speed MPC controller model for a specific vehicle while retaining cross-terrain generalization capability. The method uses FiLM to summarize recent state-action history into a single dynamics context token to condition a Transformer FKD architecture for multi-step prediction, and then specializes it by combining limited real-vehicle data with targeted synthetic trajectories generated by system identification. In closed-loop MPC experiments across three terrains and one out-of-distribution towing task, the paper reports that with only 5 minutes of real-vehicle data per terrain, OptCar is comparable to a dedicated model trained on 30 minutes of road data on roads, and performs better when terrain changes. The paper was accepted to CoRL 2026.
Report timeline
- OptCar: Specializing a Generalist FKD Model into a High-Speed MPC-Specific Model with Limited Real-Vehicle Data
The paper proposes OptCar, a method that specializes a generalist forward kinodynamic (FKD) model into a vehicle-specific model while preserving cross-terrain generalization. At its core, a FiLM mechanism summarizes recent state-action history into a single dynamics context token that conditions multi-step predictions in a Transformer FKD architecture; specialization then combines limited real-vehicle data with targeted synthetic trajectories generated by system identification.
arXiv cs.RO ·