munich2drive_edgar

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.

The ego is rendered with the real EDGAR mesh -- TUM's VW T7 Multivan digital twin -- rather than the stand-in van used in munich2drive. Same 50 base scenarios, same trajectories, same camera extrinsics; what changes is that the vehicle you see in the chase and bird's-eye views is the vehicle the footprint has always described.

• 500 clips over 50 base scenarios (500 sumo_edgar)
• 61,200 agent-replays, 79-161 agents per clip
• 12 weather / time-of-day conditions, 0 at night
• ego paths 10.64-232.85 m (median 108.96 m)

Camera provenance

The six surround-camera extrinsics (front, front_left, front_right, back, back_left, back_right) are EDGAR's real Basler short-range rig, read from edgar.description.states.urdf in TUM-AVS/edgar_digital_twin and converted from the URDF's ROS convention into CARLA's frame -- the poses hang off rear_axle_center, so CARLA x = urdf_x - 0.009, y = -urdf_y, z = urdf_z + 0.338, with the yaw and pitch handedness flipped. They are not derived from the bounding box, and must not be rescaled to one.

EDGAR's long-range forward stereo pair is not rendered: it has no counterpart in a 6-view surround layout. The chase and bird panels are visualisation only and do not correspond to sensors on the vehicle -- they are derived from the actor's bounding box. The commonroad panel is ground truth, not a camera.

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 0.006 m, p99 0.376 m, max 1.161 m
• height deviation: median 0.049 m, p99 0.322 m, max 0.383 m
• clips rendered with actors inherited from a previous run: 0; with actors surviving cleanup: 0
• clips with at least one ego collision event: 0

Known limitations

• Traffic is replayed kinematically from recorded SUMO trajectories and does not react to the ego. No clip in this set records an ego collision event, but the absence is a property of these particular replays, not evidence that the traffic would yield.
• Where SUMO placed two footprints overlapping, the render shows them interpenetrating.
• The ego mesh is the EDGAR digital twin as cooked for CARLA; it is hosted on the vehicle.lincoln.mkz_2017 blueprint id, which is cosmetic -- what spawns is EDGAR (bbox 4.962 x 2.233 x 2.089 m, TUM livery).
• Six clips -- scenario 08 ego 0, scenario 12 egos 8 and 9, scenario 14 ego 5, scenario 35 ego 22 and scenario 42 ego 10 -- were withheld from the first release because the ego's physics body lifted it off the road at a single audit step. They were re-rendered with the ego replayed kinematically: placed exactly on its trajectory every step, as all surrounding traffic always is. They now pass the audit with zero pose and height error. Their base scenarios were regenerated from the original SUMO inputs and seeds and checked against the published munich2drive trajectories: identical agent counts and types, and every ego state equal to the published 3-decimal precision.
• In the other 494 clips the ego carries a physics body, which exists only to drive the collision sensor. With the EDGAR mesh that sensor recorded no events in any clip -- including the five whose identical trajectories register hits in munich2drive with the stand-in van -- so treat ego_collision_events in this set as uninformative. For the six kinematic clips it is 0 by construction.
• All clips are daytime; unlike munich2drive this set contains no night renders.

Acknowledgements

Munich HD map from SimForge; ego vehicle and motion planner from TUM-AVS (EDGAR digital twin, Frenetix); built on CARLA, CommonRoad and SUMO. The EDGAR digital twin (TUM-AVS/edgar_digital_twin) is LGPL-3.0; its attribution and licence text travel with the cooked asset and with these renders of it. Observe the upstream licences of the HD map, simulator assets and vehicle models.