Driver Models in SUMO and Elsewhere
summary
Claims that there are 1,000 car‑following models and that many are redundant, but classic LNM/Gazis variants produce unbounded accelerations, diverge at short gaps, ignore driver adaptation and allow crashes; calibration to field data exposes parameter non‑identifiability (IDM with delta=4 forces Vmax unrealistically high unless delta is fit) and observability limits. Uses SimTD: 100 vehicles, 100 days, 1,000 drivers, 2 TB compressed and SUMO calibration to a DLR intersection, finds discrete action points dominate behavior and proposes focusing on discrete control and careful calibration (or data‑trained LLM/ML approaches).
AI-written summary · about
chapters (10)
- 0:00Driver models in SUMO
- 1:02Car following models history
- 3:51Beyond car following complexity
- 6:44Gazis-Herman-Rothery critique
- 9:35Empirical vs theoretical models
- 13:31Discrete control in behavior
- 16:01Calibrating SUMO challenges
- 19:57Observability and calibration issues
- 21:45IDM calibration correlations
- 23:30Conclusion and future
related talks
18mSUMO Simulations for Federated Learning in Communicating Autonomous Vehicles
33mHuman-Like Behavior in Autonomous Systems
42mAutomated Driving - Taking the Human Factors Researcher out of the Loop
31mDeep learning for self-driving vehicle technology
28mBen Lau, Quantitative Researcher, Hobbyist, NYC 2017
37mBuckle Up: How Waymo Uses Data to Create a Safer Driving Experience
39mFast Open-Loop and Closed-Loop Simulators for Autonomous Vehicles
59mOn Human-robot Interactions via Multimodal Intent Modeling and Reachability-based Safety Assurance