Dataset Viewer
Auto-converted to Parquet Duplicate
seq_id
stringlengths
22
36
range_id
int32
0
3
range_uid
stringlengths
24
38
frame_start
int32
0
1.84k
frame_end
int32
24
1.93k
n_frames
int32
25
90
distance_m
float64
2.72
360
theta_med
float64
1.24
24.9
n_merge_cc
int32
1
54
placement
stringclasses
2 values
n_turning
int32
2
24
frames
listlengths
25
90
grounding_md
stringlengths
1.76k
54.4k
map_png
unknown
chunk_rms
listlengths
3
11
f0
int32
0
41
f1
int32
16
89
n_meta_frames
int32
10
89
surfaces
listlengths
1
7
entities
listlengths
0
14
side_lat_min
float64
1.25
2.84
ambig_m
float64
0.6
2.84
subtasks
listlengths
1
38
subtask_draft
listlengths
subtask_metrics
stringclasses
0 values
residual_violations
listlengths
label_note
stringclasses
0 values
draft_raw
stringclasses
0 values
final_raw
stringclasses
0 values
l1_task
stringclasses
0 values
l1_through_line
stringclasses
0 values
l1_residual_violations
listlengths
l1_raw
stringclasses
0 values
label_model
stringclasses
1 value
UuThHmu0qyn8ncO21v4ylQ
1
UuThHmu0qyn8ncO21v4ylQ:1
222
264
43
129.422941
9.080912
1
anchored
3
[{"frame_idx":235,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An9e6-o1tB4A1n-RiIoM(...TRUNCATED)
"TOTAL_FRAMES (N): 24\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAADjtklEQVR4nOzddXgcdfoA8HdkXePuTaqp0ZYaLV6hWIs7HG4HxTk(...TRUNCATED)
[ 0.114, 1.3, 0.146, 0.172, 0.175 ]
19
42
24
[ "road" ]
[ "street lamp", "metal light tower", "concrete barrier", "chain link fence" ]
1.25
0.6
[{"frame_start":0,"frame_end":2,"reasoning":null,"end_condition":null,"instruction":"Continue straig(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
SKjIcEkwmOJDA6zQWdatiG
1
SKjIcEkwmOJDA6zQWdatiG:1
68
94
27
57.476886
5.4361
13
anchored
8
[{"frame_idx":96,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An-oVbjztFnApBebsdFC8(...TRUNCATED)
"TOTAL_FRAMES (N): 24\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAC6bElEQVR4nOzdd3xUVdoH8Gd6eu+99wqhhl4UKaKCCoIigl3svmt(...TRUNCATED)
[ 0.551, 0.544, 1.203 ]
2
25
24
[ "sidewalk", "pedestrian crossing", "road" ]
[ "bollard", "street lamp", "building entrance", "building", "tree" ]
1.25
0.8265
[{"frame_start":0,"frame_end":4,"reasoning":null,"end_condition":null,"instruction":"Walk straight a(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
SKjIcEkwmOJDA6zQWdatiG
2
SKjIcEkwmOJDA6zQWdatiG:2
171
197
27
47.61317
5.4361
7
anchored
7
[{"frame_idx":238,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An9sNykbfgAhWMjTzlE1(...TRUNCATED)
"TOTAL_FRAMES (N): 25\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAADce0lEQVR4nOzdd3xUVfYA8PPe9J5JMumNQCD00DtSpQiIHRQLghV(...TRUNCATED)
[ 0.436, 0.109, 0.483 ]
2
26
25
[ "paved sidewalk", "road", "tactile paving" ]
[ "vending machine", "bus shelter", "bench", "street lamp", "signpost", "shopfront", "railing" ]
1.25
0.654
[{"frame_start":0,"frame_end":7,"reasoning":null,"end_condition":null,"instruction":"Continue along (...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
whxr14dnifeqvmu1k5hnes
0
whxr14dnifeqvmu1k5hnes:0
670
699
30
88.262934
9.044424
4
anchored
3
[{"frame_idx":693,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An9daBKZ5oGI3N-mikGD(...TRUNCATED)
"TOTAL_FRAMES (N): 22\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAEAAElEQVR4nOzdd3gc1bUA8DNle1Xv1ZIsy7aKe7dxx9hgTO+YGAg(...TRUNCATED)
[ 0.101, 0.038, 2.264 ]
8
29
22
[ "asphalt road", "brick pavement" ]
[ "white sign", "green sign", "red sign", "black van", "brick wall", "yellow building", "hedge", "tree" ]
1.25
0.6
[{"frame_start":0,"frame_end":3,"reasoning":null,"end_condition":null,"instruction":"Walk straight a(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
FMlNxi-Lb1_gEEqLiP6WBg
1
FMlNxi-Lb1_gEEqLiP6WBg:1
555
644
90
265.677413
12.256563
1
uniform
5
[{"frame_idx":560,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An952DbiB-vnbWMZ60ac(...TRUNCATED)
"TOTAL_FRAMES (N): 87\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAEAAElEQVR4nOzdd3wb5f0H8O8t7WVZ8t5xnDh7b8iGQEgKYbWQMMp(...TRUNCATED)
[ 0.464, 0.105, 0.174, 0.249, 0.298, 0.208, 0.319, 0.269, 0.312, 0.143, 0.394 ]
3
89
87
[ "paved sidewalk", "road", "pedestrian crossing", "cobblestone street" ]
["yellow bollard","street lamp","traffic light","parked car","hedge","black fence","brick wall","ora(...TRUNCATED)
1.25
0.6
[{"frame_start":0,"frame_end":6,"reasoning":null,"end_condition":null,"instruction":"From the inters(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
RBqx9StkTprGZOmsXPUeuY
0
RBqx9StkTprGZOmsXPUeuY:0
18
53
36
107.143101
6.419817
1
anchored
3
[{"frame_idx":19,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An998XOnQQVhaY2WMktmj(...TRUNCATED)
"TOTAL_FRAMES (N): 34\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAC3X0lEQVR4nOzdd3xb5dUH8KMtD8myZct7xHb23mGEhA1hhkCBlrB(...TRUNCATED)
[ 0.202, 1.836, 0.256, 0.661 ]
2
35
34
[ "concrete path", "grass" ]
[ "street lamp", "hedge", "guardrail", "wooden fence" ]
1.25
0.68775
[{"frame_start":0,"frame_end":2,"reasoning":null,"end_condition":null,"instruction":"Continue straig(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
i0hprbezmvkhm2i6pp7ji3
0
i0hprbezmvkhm2i6pp7ji3:0
3
86
84
92.319259
18.107688
3
uniform
23
[{"frame_idx":5,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An__r8OvfGEp0FMJtrCPjp(...TRUNCATED)
"TOTAL_FRAMES (N): 53\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAEAAElEQVR4nOyddXgUV9fAz8ys78Z140IgSHB3h7ZIocXqFN4qRVu(...TRUNCATED)
[ 0.097, 0.057, 0.056, 0.1, 0.107, 0.368, 0.086, 1.595, 0.379, 0.257 ]
26
78
53
[ "gravel path", "dirt path", "pond", "wooden bridge" ]
[ "torii gate", "shrine", "red bridge", "stone lantern", "railing", "rock wall", "treeline" ]
1.25
0.6
[{"frame_start":0,"frame_end":6,"reasoning":null,"end_condition":null,"instruction":"Starting beside(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
i0hprbezmvkhm2i6pp7ji3
1
i0hprbezmvkhm2i6pp7ji3:1
96
131
36
30.1935
18.107688
4
uniform
20
[{"frame_idx":136,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An-xiXFz365COFxyj0HZ(...TRUNCATED)
"TOTAL_FRAMES (N): 18\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAADp2klEQVR4nOzdd3yU9f0A8M8zbo8kd5e99yAQ9gYRARXcgKit1qp(...TRUNCATED)
[ 0.634, 0.237, 0.052, 0.18 ]
18
35
18
[ "stone path", "gravel" ]
["stone monument","stone lantern","wooden fence","torii gate","wheelbarrow","orange cone","shrine bu(...TRUNCATED)
1.25
0.6
[{"frame_start":0,"frame_end":8,"reasoning":null,"end_condition":null,"instruction":"Starting on the(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
x-D19tGZrwqMiHlSxZrIRw
0
x-D19tGZrwqMiHlSxZrIRw:0
90
151
62
124.92646
21.898536
1
anchored
3
[{"frame_idx":182,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An-u9L8XiD_J6VYP15-7(...TRUNCATED)
"TOTAL_FRAMES (N): 51\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAD5SklEQVR4nOzddXwcZfoA8GdmXbObbHzj0qTRunuhpQ60FCulUKR(...TRUNCATED)
[ 0.116, 0.609, 0.662, 0.122, 0.076, 0.056, 0.126 ]
3
53
51
[ "asphalt road", "concrete curb", "manhole cover" ]
[ "metal railing", "white guardrail", "stone retaining wall" ]
1.25
0.6
[{"frame_start":0,"frame_end":3,"reasoning":null,"end_condition":null,"instruction":"Walk straight a(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
x-D19tGZrwqMiHlSxZrIRw
1
x-D19tGZrwqMiHlSxZrIRw:1
172
212
41
49.259947
21.898536
1
anchored
11
[{"frame_idx":321,"image":{"path":"https://scontent-sjc6-1.xx.fbcdn.net/m1/v/t6/An8rvQN5bt8OXIARQRp3(...TRUNCATED)
"TOTAL_FRAMES (N): 34\n\nPATH (per frame: heading change entering the frame, + = bends right; hop le(...TRUNCATED)
"iVBORw0KGgoAAAANSUhEUgAAApQAAAM5CAIAAAA2WDbvAAEAAElEQVR4nOzdZXwU19oA8GfWPdlsXAkRQgghWCAEdynSUqC0tKU(...TRUNCATED)
[ 0.493, 0.214, 0.084, 0.12, 0.548 ]
7
40
34
[ "road", "pedestrian crossing" ]
[ "manhole cover", "information board", "metal railing", "stone wall", "concrete wall" ]
1.25
0.6
[{"frame_start":0,"frame_end":5,"reasoning":null,"end_condition":null,"instruction":"Follow the side(...TRUNCATED)
null
null
null
null
null
null
null
null
null
null
Qwen/Qwen3.8-Flash-Next-FP8
End of preview. Expand in Data Studio

navworld_extra_relabeled

Additional street-level navigation episodes from Mapillary street-level image sequences, built the same way and with the same schema as cedarfield482/navworld_49cities_relabeled. It is meant to be added to that dataset's training split.

One row is one range: a continuous walk or ride through a street scene, on average 49 frames and 117 m long. A row is self-contained. It carries the images, the metric camera pose and motion of every frame, a grounded description of the scene, and an ordered list of short navigation instructions ("sub-tasks") that together cover the whole range.

Split Shards Ranges Frames Distance Sub-task instructions
train 2,106 19,216 947,373 2,249 km 268,269

This repository has a train split only.

Source and licence

The images come from Mapillary, which publishes street-level imagery under CC BY-SA 4.0. The images were processed (rectified and re-encoded) for this dataset. seq_id is the Mapillary sequence each range was taken from, and frames[*].image.path is the address each image was downloaded from. Per-image contributor credit is not included yet.

This dataset, including the poses, scene grounding and instructions added to the images, is released under CC BY-SA 4.0.

Loading

import io
from datasets import load_dataset
from PIL import Image

ds = load_dataset("cedarfield482/navworld_extra_relabeled", split="train", streaming=True)
row = next(iter(ds))
image = Image.open(io.BytesIO(row["frames"][0]["image"]["bytes"]))
for st in row["subtasks"]:
    print(st["frame_start"] + row["f0"], st["frame_end"] + row["f0"], st["instruction"])

Schema

34 columns per row.

Range

Column Type Meaning
seq_id string Source image sequence.
range_id int32 Index of the range within its sequence.
range_uid string "<seq_id>:<range_id>", unique per row.
frame_start, frame_end int32 Position of the range in the source sequence, inclusive.
n_frames int32 Number of frames, equal to len(frames).
distance_m double Path length of the range in metres.
theta_med double Median angle between camera direction and travel direction for the sequence (degrees).
n_merge_cc int32 Number of reconstruction components the range spans.
placement string How the range was placed: anchored (centred on a turn) or uniform.
n_turning int32 Number of turning frames in the range.

frames (list of structs, one per frame, in travel order)

Field Type Meaning
frame_idx int32 Position of the frame in the source sequence.
image struct bytes: the encoded image; path: the source URL it came from.
cam2world double[16] 4x4 camera-to-world pose, row-major, in a metric frame local to the sequence.
camera_type, camera_parameters string, list Camera model and its parameters.
captured_at int64 Capture time, milliseconds since epoch.
height, width int32 Size of the original image.
is_pano bool Whether the source image is a panorama.
merge_cc int64 Reconstruction component of the frame.
dt_cam double[3] Translation to the next frame, in this frame's camera coordinates (metres). NaN on the last frame.
dyaw_deg double Yaw change to the next frame (degrees). NaN on the last frame.
hop_m double Horizontal distance to the next frame (metres). NaN on the last frame.
path_turn_deg double Bend of the route at this frame (degrees, positive = right).
cam_offset_deg double Angle between camera direction and travel direction (0 to 180 degrees); null when not moving.
motion_category string STATIONARY, FORWARD, TURNING_LEFT, TURNING_RIGHT, BACKWARD or LATERAL.

The actions of a range are the motions of frames[:-1]; the last frame has no successor inside the range.

Scene grounding

Column Type Meaning
grounding_md string Text description of the scene from 3D grounding: the path, surfaces, objects and which side they are on.
map_png binary Rendered top-down map of the scene (PNG).
f0, f1 int32 First and last frame of the grounded window, as inclusive positions into frames.
n_meta_frames int32 f1 - f0 + 1; grounding_md numbers its frames 1 to n_meta_frames.
surfaces, entities list of string Ground surfaces and objects found in the scene.
chunk_rms list of double Alignment error of the 3D reconstruction per chunk (metres).
side_lat_min, ambig_m double Distance thresholds (metres) used to decide left/right and to flag ambiguous positions.

Sub-tasks

Column Type Meaning
subtasks list of struct Ordered instructions. Each has frame_start, frame_end and instruction; reasoning and end_condition are null.
label_model string Model that wrote the instructions.

Sub-task frame indices are 0-based inside the grounded window: the frame in frames is frames[frame_start + f0]. Consecutive sub-tasks share their boundary frame (frame_end of one equals frame_start of the next), and every range is covered from start to end. Boundaries are placed from the route geometry, so that a turn is never split and a leg is roughly 3 to 10 m long; each instruction is then written by a vision-language model from the frames of its leg and the measured distance and turn.

Unused columns

These columns are part of the schema and are null in every row: subtask_draft, subtask_metrics, residual_violations, label_note, draft_raw, final_raw, l1_task, l1_through_line, l1_residual_violations, l1_raw.

Downloads last month
63