---
pretty_name: gt-traces
tags:
- robotics
- navigation
- legged-robot
- anymal
- grand-tour
- trajectory
- vqa
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: mission
    dtype: string
  - name: frame_idx
    dtype: int64
  - name: image
    dtype: image
  - name: image_overlay
    dtype: image
  - name: image_width
    dtype: int64
  - name: image_height
    dtype: int64
  - name: polyline_xy
    list:
      list: float64
  - name: surface_seq
    list: string
  - name: transitions
    list:
    - name: from_surface
      dtype: string
    - name: to_surface
      dtype: string
    - name: x
      dtype: float64
    - name: y
      dtype: float64
    - name: waypoint_start
      dtype: int64
    - name: waypoint_end
      dtype: int64
  - name: offground_frac
    dtype: float64
  - name: usable
    dtype: bool
  - name: z_profile
    list: float64
  - name: surface_runs
    dtype: string
  - name: mean_luma
    dtype: float64
  - name: lighting
    dtype: string
  - name: snowy
    dtype: bool
  - name: objects
    dtype: string
  splits:
  - name: train
    num_bytes: 7205706958
    num_examples: 16421
  download_size: 7201013199
  dataset_size: 7205706958
---

# gt-traces

GrandTour (ANYmal legged robot) surface-trace dataset: first-person frames
with the future *driven* path projected into the image, plus every derived
label precomputed as columns. One row = one frame; clean and overlay
renderings both embedded. All labels come from the robot's own logs and
automatic segmentation — no human annotation.

Built from the [GrandTour](https://grand-tour.leggedrobotics.com/) ANYmal
missions (`hdr_front` camera, DLIO map odometry), 48 missions.

## Summary

- Total samples: **16,421** (48 missions)
- Named surface transitions: **2,840** across **2,247** rows
- `usable` (surface labeling gate): **13,202** (80.4%)
- Conditions: **763** snow-covered rows; lighting **14,614** daylight /
  **1,486** dusk / **321** night

## Schema

| Field | Type | Notes |
|---|---|---|
| `mission` | string | GrandTour mission name (encodes local start time) |
| `frame_idx` | int | per-mission frame index |
| `image` | Image | clean hdr_front frame (max side 1024) |
| `image_overlay` | Image | same frame with the future driven path drawn |
| `image_width` / `image_height` | int | pixel size (polyline coordinate space) |
| `polyline_xy` | List[[x, y]] | driven path projected into the image, pixel coords, ≤60 waypoints at 0.25 m arc steps |
| `surface_seq` | List[string] | canonical ground-surface label per waypoint (Mask2Former-COCO-panoptic; same merged vocabulary as `adipotnis/ts-polyline-v2`, plus `metal grate` from mission context) |
| `transitions` | List[struct] | named surface crossings `{from_surface, to_surface, x, y, waypoint_start, waypoint_end}`, x/y normalized |
| `offground_frac` | float | fraction of waypoints on non-ground classes |
| `usable` | bool | `offground_frac <= 0.5` |
| `z_profile` | List[float] | ground height per waypoint from odometry; `[]` where unavailable |
| `surface_runs` | string | JSON `[{label, start, end, verb}]` — surface runs with `walks on` / `climbs` / `descends` verbs (±0.25 m z change over a run) |
| `mean_luma` | float | mean grayscale value of the frame |
| `lighting` | string | daylight / dusk / night (mission start hour + luma override) |
| `snowy` | bool | any `snow` waypoint label in the frame |

## Label provenance

- **Path**: the robot's actually-driven trajectory (DLIO odometry) — self-labeled.
- **Surfaces/transitions**: `facebook/mask2former-swin-large-coco-panoptic`,
  canonical word-matched vocabulary shared with the wheeled dataset; crossings
  need ≥3-waypoint runs each side, narrow unknown gaps (≤5 waypoints) bridged.
- **Climb/descend verbs**: z change across a surface run from the elevation
  profile.
- Build script: `ts-polyline-pipeline/grand_tour/build_gt_traces.py`.

Access is manually gated — request access and briefly say what for.

If you use this dataset, please also credit the GrandTour dataset
(Fankhauser et al., ETH Zürich Robotic Systems Lab).