ts-polyline

TerraSentia (wheeled robot) navigation trajectories from ROS bags, packaged in a schema mirroring adipotnis/scand-polyline. Each row is one image frame paired with a 60-waypoint future trajectory (next 15 m of EKF odom sampled at 0.25 m steps), projected into the ZED camera as a pixel-space polyline and rendered as an overlay PNG.

Summary

• Total samples: 108491
• Bags: 22
• Frames flagged collision: 16372 (15.1%)
• Frames flagged reversing: 26300 (24.2%)
• Distinct collision events across all bags: 244

Schema

Field	Type	Notes
traj_id	string	bag stem
robot	string	constant "Wheeled Robot"
run_id	int32	per-bag index
sub_id	int32	0 (schema parity with scand-polyline)
n_frames	int32	total frames in this bag
frame_idx	int32	per-bag frame index
prefix	string	same as traj_id
image	Image	rectified ZED left frame
image_overlay	Image	same frame with projected polyline drawn
depth	Image	nearest ZED registered depth frame, 16-bit PNG in millimeters (0 = invalid). Decode: np.asarray(row['depth']).astype(np.float32) / 1000.0 to get meters.
traversability	Image	nearest /terrasentia/traversability_map frame, mono8 PNG, 336x188 (lower res than RGB).
embodiments	List[string]	["Wheeled Robot"]
polyline_xy	List[List[float64,2]]	projected pixel polyline (in-image segment)
traj	List[List[float64,3]]	up to 60 base-frame waypoints (next 15 m @ 0.25 m)
ground_truth	dict	{"Wheeled Robot": [polyline_xy]}
image_width / image_height	int64	rectified frame dims
collision	bool	a confirmed multi-signal collision event occurs in the next 5 s
time_to_collision	float32	seconds until that event; -1.0 when collision=False
reversing	bool	sustained reverse motion in next 5 s

Per-bag top-view maps

Forward motion in green, reverse motion in blue, detected collision events as red dots.

ts_2023_04_10_20h05m59s_filtered

• frames: 4334 · collision frames: 285 · reversing frames: 1502 · collision events: 4

ts_2023_04_10_20h12m46s_filtered

• frames: 4508 · collision frames: 150 · reversing frames: 1540 · collision events: 2

ts_2023_04_11_17h47m24s_filtered

• frames: 2276 · collision frames: 258 · reversing frames: 570 · collision events: 4

ts_2023_04_11_17h51m03s

• frames: 6077 · collision frames: 582 · reversing frames: 1000 · collision events: 8

ts_2023_04_11_17h53m56s_filtered

• frames: 6478 · collision frames: 1298 · reversing frames: 1589 · collision events: 21

ts_2023_04_11_17h58m32s

• frames: 9979 · collision frames: 524 · reversing frames: 1464 · collision events: 7

ts_2023_04_11_18h00m20s

• frames: 7276 · collision frames: 944 · reversing frames: 1781 · collision events: 15

ts_2023_04_11_18h02m48s_filtered

• frames: 6799 · collision frames: 881 · reversing frames: 1804 · collision events: 15

ts_2023_04_11_18h13m59s

• frames: 2983 · collision frames: 224 · reversing frames: 443 · collision events: 3

ts_2023_04_11_19h57m53s

• frames: 3693 · collision frames: 75 · reversing frames: 1037 · collision events: 1

ts_2023_04_11_20h02m40s

• frames: 2022 · collision frames: 75 · reversing frames: 502 · collision events: 1

ts_2023_06_25_20h32m20s

• frames: 6417 · collision frames: 1105 · reversing frames: 1904 · collision events: 17

ts_2023_06_25_23h24m19s

• frames: 5280 · collision frames: 1097 · reversing frames: 1403 · collision events: 16

ts_2023_06_25_23h34m37s

• frames: 7269 · collision frames: 1004 · reversing frames: 1542 · collision events: 14

ts_2023_06_25_23h56m10s

• frames: 2050 · collision frames: 862 · reversing frames: 319 · collision events: 17

ts_2023_06_25_23h56m47s

• frames: 6002 · collision frames: 592 · reversing frames: 1470 · collision events: 8

ts_2023_06_26_00h04m16s

• frames: 973 · collision frames: 74 · reversing frames: 134 · collision events: 1

ts_2023_06_26_00h05m52s

• frames: 2383 · collision frames: 493 · reversing frames: 394 · collision events: 8

ts_2023_06_26_00h05m58s

• frames: 4621 · collision frames: 766 · reversing frames: 1440 · collision events: 12

ts_2023_06_26_00h14m59s

• frames: 1989 · collision frames: 552 · reversing frames: 549 · collision events: 8

ts_2023_06_26_00h15m21s

• frames: 7259 · collision frames: 2093 · reversing frames: 1939 · collision events: 29

ts_2023_06_26_00h25m04s

• frames: 7823 · collision frames: 2438 · reversing frames: 1974 · collision events: 33

Detection heuristics

Collision events are detected once per bag by fusing four streams (datasets/ts_data/scripts/collision_events.py), then each frame is labeled by whether a confirmed event falls in the 5 s after it. All EKF timing uses msg.header.stamp — the rosbag receipt timestamps for /terrasentia/ekf are batched in microseconds and don't reflect the true ~100 Hz sample times.

Candidates (from odometry, required):

• sharp decel: rolling-median-smoothed (0.3 s) EKF speed drops from > 0.4 m/s to < 0.12 m/s within 0.5 s. Raw finite-difference speed has single-sample glitches up to 15 m/s, so smoothing is essential — the previous raw-speed rule flagged ~50 % of all frames.
• stuck: forward motion commanded (/terrasentia/motion_command vx > 0.3 m/s) or wheels driving forward (signed mean of /terrasentia/motors wheel speeds > 0.3 m/s) while the body moves < 0.08 m/s for >= 0.5 s, excluding in-place turns (|yaw rate| > 0.3 rad/s). Catches low-speed bumps and wheel-slip pushes that never produce a sharp decel.

Corroboration (at least one required to confirm):

• depth proximity: in a lower-center ROI of the nearest ZED depth frame, > 20 % of pixels closer than 0.75 m, or the invalid-pixel fraction surges > 0.25 above the rolling scene baseline (objects closer than the ZED2 min range — including anything within ~6 in — decode as invalid, so a surge in invalid pixels is a proximity signal).

• IMU contact spike: jerk (|Δa|/Δt) exceeds 10x the bag's median within ±0.6 s of the candidate; also refines the event timestamp.

• commanded-vs-actual mismatch: forward commanded (stick or wheels) while the body is stationary near the candidate.

• reversing: smoothed signed velocity along robot heading is < -0.05 m/s for more than 30 % of samples in the window.

Built with datasets/ts_data/scripts/bags_to_hf_polyline.py.