Yangyihui/ds-be75b4e409e549321871

Sequence Object Navigation R5

This is a non-commercial research dataset derived from SpatialVID-HQ. The repository identifier is deliberately content-neutral, while this card documents the contents, provenance, filtering policy, and license explicitly.

It contains 13,647 complete egocentric RGB videos and 18,249 object-goal navigation windows. Every instruction has the exact form Go to <object>. and the selected object is intended to be visible in the first frame. Videos are H.264/yuv420p at constant 4 FPS and 512 x 288. Segment ranges use exclusive-end [t0, t_end) indexing and every window contains exactly 9*n frames.

Layout

data/p21-00000.tar
data/p21-00001.tar
...
metadata/index.jsonl
metadata/segments.jsonl
metadata/dataset_summary.json
metadata/release_audit.json
metadata/package_audit.json
metadata/shard_manifest.json
metadata/SHA256SUMS.shards.jsonl

Each uncompressed tar contains paired files:

videos/<clip>.mp4
labels/<clip>.json

Each label is a list because one source route may provide several overlapping navigation windows. Labels include the instruction, object description, frame range, first-frame target box in 0-1000 coordinates, arrival frame, geometric motion summaries, selected official SpatialVID text annotations, and review tags.

Construction and quality control

Candidate endpoints were discovered from annotated camera poses, target objects were selected with the official scene text as auxiliary context, and valid starts were searched backward in exact 9-frame chunks. The generation pipeline checked first-frame visibility, target grounding, meaningful translation, goal-relative progress, arrival, horizontal/vertical motion, and route geometry.

A final dense-frame review used one Gemini 3.1 Flash Lite call per accepted segment. Only confident, material nonphysical overlays such as burned-in subtitles, editor titles, watermarks, or UI caused hard removal. Navigation concerns were retained as soft tags so downstream users can choose their own threshold. This removed 411 source videos and 642 candidate segments:

All labels and videos were checked for schema, timing, frame-range, resolution, frame-rate, and host-path leakage. All videos were probed, all newly encoded videos were decoded by FFmpeg during transcoding, reused videos came from a prior full decode audit, and a 256-video stratified sample was decoded again. Every release tar was reopened and verified against its SHA-256 digest before upload.

Intended use

This release is intended for action-free video pretraining, instruction-conditioned video modeling, visual navigation representation learning, and related research. It contains no robot action commands, odometry, depth, or camera calibration and must not be treated as executable Go2 control data.

Limitations

• —Instructions and reviews are machine generated and can still contain semantic, grounding, timing, or arrival errors.
• —Videos are in-the-wild camera routes rather than trajectories generated under quadruped kinematic, collision, or embodiment constraints.
• —Soft-tagged navigation mismatches remain in the release by design.
• —Physical scene text, people, brands, and other source-video content may remain.

License and attribution

This derived release uses CC BY-NC-SA 4.0, matching the upstream ShareAlike requirement. Use is non-commercial; credit SpatialVID-HQ and distribute adaptations under the same license. See the upstream dataset card and license terms.