iprlnav3/nav3_500k_cam
nav3 training data (hive pipeline) Assembled from the hive-format V2 pipeline (01→11→03→04→05→filter_chunks_ray→ label_chunks_ray). One row per priority-frame-anchored GOOD chunk: the chunk's frames (images_rgb), the per-frame (dx, dz, dyaw_deg) actions (+ bucketized action_tokens using the action_bins.json sidecar, n_bins=16), the full L1–L8 instruction tree from the local-GPU VLM labeler, and camera geometry: cam2world — (n_frames, 4, 4) per-frame extrinsics (OpenCV… See the full description on the dataset page: https://huggingface.co/datasets/iprlnav3/nav3_500k_cam.
nav3 training data (hive pipeline)
Assembled from the hive-format V2 pipeline (01→11→03→04→05→filterchunksray→ labelchunksray). One row per priority-frame-anchored GOOD chunk: the chunk's frames (images_rgb), the per-frame (dx, dz, dyaw_deg) actions (+ bucketized action_tokens using the action_bins.json sidecar, n_bins=16), the full L1–L8 instruction tree from the local-GPU VLM labeler, and camera geometry:
cam2world—(n_frames, 4, 4)per-frame extrinsics (OpenCV cam→local-world; origin = sequence frame 0, +X East, +Y Down, +Z North).camera_type/camera_parameters— Mapillary intrinsics (normalized[focal, k1, k2], focal = f / max(w, h) of the ORIGINAL capture); constant within a chunk.camera_parametersis empty (andcamera_typenull) for the rare chunk whose source metadata is missing.
priority_frame_idx (chunk-local) splits history/future for images_rgb, actions, and cam2world. Load action_bins.json via huggingface_hub.hf_hub_download(repo_id, "action_bins.json", repo_type="dataset").