This dataset supports research on learning driver behavior in real-life settings, with 104 hours of human driving data from the San Francisco Bay Area. It includes sensor readings, GPS data, and front-facing camera footage.
Challenging dataset to enable research on learning driver behavior in real-life environments. The dataset includes 104 hours of real human driving in the San Francisco Bay Area collected using an instrumented vehicle equipped with different sensors. The videos below provide examples of goal-oriented driving behavior from the dataset. GPS coordinates together with sensor values from the CAN bus are shown on top of the front-facing camera stream
Use case:
Learning Human Driver Behavior from Real-World Multi-Sensor Driving Data
| Example Observation vector (oₜ) ≈ 80 dims | – ego-state : speed, accel, yaw-rate, steering-angle, gear, engine-RPM, fuel-rate – environment : lead-car distance & Δv (from LiDAR), lane curvature, traffic-light state, speed-limit, road grade, GPS coord. – context : time-of-day, weather tag |
| Example Action space (aₜ) | • continuous 3-d: – throttle ∈ [0, 1] – brake-pressure ∈ [0, 1] – steering-wheel rate ∈ [-0.5, 0.5] rad/s |
| Example Reward rₜ | rₜ = − 0.05 · fuel_flow(g/s) (efficiency) − 0.1 · |jerk| (comfort) − 1 · 1[collision] (safety, sparse) − 0.02 · 1[|v − v_limit|>2 m/s] (compliance) |
| Example Transitions | • 104 h × 3600 s/h × 10 Hz ≈ ≈ 3.7 M (s, a, r, s′) tuples. • Empirical action-change probability ≈ 0.36 per step (drivers hold pedal for ~0.28 s ⇒ moderate temporal correlation). |