dreamer_lc_v2

Regenerated lane-change counterfactuals and lane-change action labels for the SimLingo CARLA dataset. The files only store per-frame labels and trajectories. You still need the SimLingo data itself (data/simlingo/...) for images and measurements.

dreamer_lc_v2.tar.gz contains three NumPy .npz files:

File	Size	Content
lane_change_v2.npz	133 MB	All regenerated lane-change candidates (1,197,005) for 2,061,176 frames on 16,255 routes, with every check field
cf_lane_change_v2.npz	12 MB	Training subset: class-1 lane changes (116,471 candidates) and class-2 "wait for a gap" frames (383,756)
bypass_index.npz	32 KB	Class-4 frames: real expert bypasses of construction sites / parked vehicles (1,329 segments, 39,188 frames)

Conventions

• A frame is (route_names[route_id], frame). The route name is the route directory relative to <data_path>/data/simlingo/, and frame is the 4 Hz frame number (measurements/{frame:04d}.json.gz, rgb/{frame:04d}.jpg).
• Coordinates are in the ego frame of the current frame, in meters: x forward, y right.
• waypoints (10 × 2) are the ego positions at 0.25 (k+1) s, k = 0..9. This matches the expert waypoint labels of the SimLingo dataloader.
• route (20 × 2) is the shifted route, sampled every 1 m from the ego position.
• side: 0 = left, 1 = right. type: 1 = same-direction lane, 2 = oncoming lane.
• Only frames the SimLingo dataloader uses are covered: routes whose results.json.gz status is ok, and frames in [10, num_seq − 13). Routes under other_scenario_* are excluded.

How the lane changes were generated

The candidates replace the original SimLingo "dreamer" lane-change candidates for the adjacent lane (left or right), when that lane is a driving lane in the same direction or an oncoming lane. The generator does not need CARLA.

• Lateral motion: starts at once (start offset 0). The shift follows a smoothstep curve over max(5 m, speed × T) of distance, with T drawn uniformly from 0.8–1.5 s. The shifted route is long enough for the full 2.5 s horizon.
• Speed: the expert's real speed profile from the log. A kinematic bicycle model with the original lateral PID only controls heading and position.
• Other vehicles: their real future boxes from the log, including vehicles that only enter the sensor range during the 2.5 s.
• Skip conditions: frames are skipped as in the original generator: in a junction or within 10 m of one, route changes in the next frames, a lane-change command, or the ego in a parking lane.

lane_change_v2.npz

Per frame (2,061,176 rows):

Field	Meaning
frame_route_id, frame_frame	Frame key
frame_status	0 = evaluated, 1 = skipped by the generator, 2 = missing future frames

Per candidate (1,197,005 rows):

Field	Meaning
route_id, frame, side, type	Frame key, side and lane type
waypoints, route	Trajectory and shifted route
width	Lane width (lateral shift), m
trans, T	Transition length (m) and time (s)
crash, crash_t0, lat_at_crash	Oriented-box collision with a logged vehicle or pedestrian within 2.5 s, first collision step, and lateral offset at that step
alongside_t0	A vehicle in the target lane right beside the ego at t = 0
tgt_gap	Minimum bumper gap to target-lane vehicles while the ego overlaps them laterally (99 = none)
ttc_end	After 2.5 s, both the ego and target-lane vehicles are extrapolated at constant speed along the lane: time until they would meet, in s (99 = never)
dist_junction	Distance to the next junction along the route, m (NaN = none)
disp	Path length of the 2.5 s trajectory, m
lat_end, lat_max	Final / maximum lateral offset towards the target side, ego frame
lat_end_route, lat_max_route	Same, measured relative to the original route (removes road curvature)
track_err_max	Maximum distance between the trajectory and the shifted route
speed	Ego speed at t = 0, m/s

Classes in cf_lane_change_v2.npz

Both classes use the default thresholds of classify_v2 and classify_wait_v2.

Class 1, "Change lane to the left/right." (lc_* arrays). A candidate qualifies when all of these hold:

• the lane is at least 2.5 m wide;
• no collision, and the ego moves at least 5 m in 2.5 s;
• the final lateral shift relative to the original route is 0.8–1.2 lane widths, and the tracking error is at most 0.5 m;
• no vehicle alongside at t = 0;
• tgt_gap is at least max(2 m, 0.5 s × speed), and ttc_end is at least 2 s;
• the junction is far enough away. For the oncoming lane, there is no junction within the 2.5 s travel + 2 s × speed + 10 m. For a same-direction lane, the ego does not reach a junction within 2.5 s.

Class 2, "Wait for a gap before changing lanes." (wait_* arrays). A frame qualifies when every usable adjacent lane is blocked by target-lane traffic and no candidate qualifies for class 1.

• A lane is blocked when the ego moves and at least one of these holds: a vehicle alongside at t = 0, the gap is too small, ttc_end < 2 s, or a collision after the ego has moved into the target lane (lateral offset ≥ 1.5 m).
• Frames where a candidate collides while still in its own lane (rear-ending the vehicle ahead) are excluded.
• wait_crash marks frames where some candidate collides in the target lane.

Counts (frames, after the dataloader's filters):

Split	Class 1	Class 2
Town12 + old towns (routes_training)	61,633	199,516
Town13 (routes_validation)	36,056	166,837
parking-lane routes	2,238	17,403

bypass_index.npz

Frames from real expert bypasses of ConstructionObstacle(TwoWays) and ParkedObstacle(TwoWays). The answer comes from the SimLingo commentary stage of each frame, and the trajectory is the expert's own.

• Segments: a bypass segment is the set of consecutive frames where the planned route deviates laterally by at least 2 m from the original route. The scenario is the majority scenario_name in the commentary.
• Answers:
  • commentary "Wait for a gap …" → wait;
  • commentary "… as the gap is big enough" with at least 0.5 m of lateral progress in the next 2.5 s → lane_left / lane_right;
  • the same commentary without lateral progress (braking before a wait) is dropped.
• Dropped segments and frames:
  • segments that switch between wait and lane change 3 or more times;
  • same-direction wait frames whose cause is not visible in the image (cause_object_visible_in_image is False).

Field	Meaning
route_names, scenario_names, answer_names	Lookup tables
seg_route_id, seg_scenario, seg_side	One row per segment
frame_seg, frame, answer, visible	One row per frame: segment, frame number, answer index, cause visible in the image

Loading

import numpy as np

z = np.load("cf_lane_change_v2.npz")
names = z["route_names"]
i = 0
route_rel, frame = names[z["lc_route_id"][i]], int(z["lc_frame"][i])
side = ("left", "right")[z["lc_side"][i]]
waypoints = z["lc_waypoints"][i]   # (10, 2), ego frame, 0.25 s apart
route = z["lc_route"][i]           # (20, 2), 1 m apart