| --- |
| pretty_name: ts-polyline-v2 |
| tags: |
| - robotics |
| - navigation |
| - terrasentia |
| - polyline |
| - trajectory |
| - vqa |
| --- |
| |
| # ts-polyline-v2 |
|
|
| [`adipotnis/ts-polyline`](https://huggingface.co/datasets/adipotnis/ts-polyline) |
| with every derived label baked in as columns — no sidecar files, no |
| training-time joins. One row = one camera frame with its future driven |
| trajectory and **all** supervision signals side by side. |
|
|
| TerraSentia (wheeled field robot) navigation from 22 ROS bags. Per frame: the |
| next 15 m of EKF odometry sampled at 0.25 m steps (up to 60 waypoints), |
| projected into the ZED camera as a pixel-space polyline and rendered as an |
| overlay. Every label is derived from the robot's own logs — no human |
| annotation. |
|
|
| ## Summary |
|
|
| - Total samples: **108,491** (22 bags) |
| - `collision`: **16,372** (15.1%) · `reversing`: **26,300** (24.2%) |
| - Named surface transitions: **9,655** across **8,269** rows |
| - `usable` (surface labeling gate): **71,050** (65.5%); **10,793** rows |
| geometry-rejected (polyline projects into an obstacle within 1 m — |
| wall-projection artifacts) |
| - Wheel-slip coverage: **99%** of frames scoreable; traction mean **90.3** |
| (p10 82, p90 98, 0–100 scale) |
| - Object descriptions: **20,000** frames |
|
|
| ## Schema |
|
|
| All v1 columns are unchanged (`image`, `image_overlay`, `depth` [16-bit PNG, |
| mm], `traversability`, `polyline_xy`, `traj`, `collision`, |
| `time_to_collision`, `reversing`, ids). New columns: |
|
|
| | Field | Type | Notes | |
| |---|---|---| |
| | `surface_seq` | List[string] | canonical ground-surface label per waypoint (Mask2Former-COCO-panoptic, synonym-merged vocabulary: pavement/sidewalk/path/road are ONE class; `offground` = non-ground) | |
| | `transitions` | List[struct] | named surface crossings along the path: `{from_surface, to_surface, x, y, waypoint_start, waypoint_end}`, x/y normalized image coords | |
| | `offground_frac` | float | fraction of waypoints on non-ground classes | |
| | `geom_ok` | bool | False = the drawn path enters an obstacle within 1 m arc (depth-walk test) — projection artifact, exclude from path tasks | |
| | `collision_arc_m` | float | arc length to the first waypoint whose scene depth is shorter than its forward range; −1 if the path is clear | |
| | `collision_point` | List[float] | `[x, y]` normalized coords of the blocking pixel; `[]` if clear | |
| | `usable` | bool | `geom_ok and offground_frac <= 0.5` | |
| | `slip_seq` | List[float] | wheel slip in [0,1] at the moment the robot actually crossed each waypoint: `1 − v_EKF / mean(|wheel linear_speed|)` from `/terrasentia/motors` vs EKF; −1 = no valid wheel sample (wheel speed < 0.3 m/s or > 0.5 s gap). Traction = 1 − slip | |
| | `objects` | string | JSON list of visible objects `{label, cx, cy, bin}` (normalized center, near/mid/far distance band) from SAM3 detections + depth; `""` where not computed | |
|
|
| ## Label provenance |
|
|
| - **Surfaces/transitions**: `facebook/mask2former-swin-large-coco-panoptic` |
| semantic segmentation, canonical word-matched vocabulary; crossings need |
| ≥3-waypoint runs each side, narrow unknown gaps (≤5 waypoints) bridged. |
| - **Collision geometry**: per-waypoint comparison of measured scene depth vs |
| the waypoint's robot-frame forward range. |
| - **Slip**: two independent logs (wheel encoders vs EKF body speed) sampled |
| at each waypoint's interpolated traversal time — *experienced* traction, |
| complementing the camera-predicted `traversability` map. |
| - Build scripts: `ts-polyline-pipeline` (`scripts/build_ts_polyline_v2.py` |
| and the sidecar builders it consumes). |
|
|
| Access is manually gated — request access and briefly say what for. |
|
|
