munich2drive

Surround-camera driving clips from a real Munich HD map built by SimForge, each paired with the ego vehicle's full trajectory. Every clip is a nine-panel video -- six surround cameras at the ego's real sensor extrinsics, a chase view, a bird's-eye view, and the frame-locked CommonRoad ground truth -- so what the model sees and what the data says are visible side by side.

• 500 clips over 50 base scenarios (500 sumo)
• 61,200 agent-replays, 79-161 agents per clip
• 12 weather / time-of-day conditions, 20 at night
• ego paths 10.64-232.85 m (median 108.96 m)

Each row

field
mp4	the 3x3 grid clip (1920x1080, 10 fps, 120 frames)
json	everything below, as one record
trajectory	parallel arrays: t, x, y, carla_x, carla_y, yaw, v, a, yaw_rate
path_length_m, speed_mps	trajectory summary
weather, sun_altitude_deg, precipitation, wetness, night	conditions
n_agents, agents_by_type	surrounding traffic
ego_collision_events, ego_collisions	ego collision-sensor events with impulses
audit_clean, max_pose_error_m, max_height_dev_m	per-clip verification, below

x/y are the CommonRoad map frame -- the frame the CommonRoad panel of the video is drawn in. carla_x/carla_y are the same states in the CARLA world frame, via the verified transform (carla_x = lx, carla_y = -ly after reprojecting EPSG:3857 to the map's local transverse-Mercator frame; crdesigner emits 3857, which is ~1.5x scale-distorted at 48N, so this conversion is not optional).

The trajectory is given in world coordinates; ego-relative future waypoints follow from the state at any t by the usual rigid transform.

Verification

Rendering audits each clip against its scenario file at 14 time steps and ships the result, so these are measurements, not claims:

• audit_clean: 500 / 500
• agent pose error vs CommonRoad: median 0.003 m, p99 0.022 m, max 1.297 m
• height deviation: median 0.034 m, p99 0.248 m, max 0.302 m
• clips rendered with actors inherited from a previous run: 0; with actors surviving cleanup: 0
• clips with at least one ego collision event: 6 of 500. Of the 11 events, 7 are contacts with static traffic signs and 4 with vehicles -- the sensor reports contact with anything, so this is not a vehicle-collision rate

Known limitations

• Traffic is replayed kinematically from recorded SUMO trajectories and does not react to the ego. The ego collisions present are therefore predominantly rear-end contacts caused by non-reactive traffic, not ego-at-fault conflicts.
• Where SUMO placed two footprints overlapping, the render shows them interpenetrating.
• The ego is rendered with a stand-in mesh (vehicle.volkswagen.t2_2021), identical across all clips. Its footprint throughout the CommonRoad and SUMO data is the real EDGAR size, 4.973 x 1.941 x 2.320 m, so the trajectories are unaffected.

Acknowledgements

Munich HD map from SimForge; ego vehicle and motion planner from TUM-AVS (EDGAR digital twin, Frenetix); built on CARLA, CommonRoad and SUMO. Observe the upstream licences of the HD map, simulator assets and vehicle models.