anhdao69/S8N-dagger-data

S8N DAgger: R2R and English-guide RxR 15 degrees

Collected with the pinned S8N epoch-1 policy using Janus-style recovery. Both datasets use 15-degree turns, 0.25 m forward steps and lossless 640x480 RGB. Only accepted, completed recovery trajectories are training examples. The audit covers all original training IDs, including rejected attempts and reviewed replacements. No validation/test split or scene mesh is included.

Layout and loading

R2R/labels/train/train.json.gz and RxR_15deg/labels/train/train_guide.json.gz use the existing VLN-CE episode format. Matching _gt.json.gz files contain the collected executed actions and forward locations, not the original oracle rollout. Instructions, starts, goals and reference paths are retained from the source labels. Missing RxR timed-word start/end timestamps represented upstream as NaN become JSON null, with each change listed in audit/metadata_normalizations.json. Instruction text and geometry are unchanged. All frames are captured before their named action. Actions are 0=STOP, 1=MOVE_FORWARD, 2=TURN_LEFT, 3=TURN_RIGHT. There is exactly one real final STOP; no leading sentinel or synthetic final label is added.

Extract every independent R2R/images-*.tar and RxR_15deg/images-*.tar into the dataset root. They contain <dataset>/train/<episode_id>/step_000000_ACTION.png. The existing SimpleMemVLN prepare_r2r.py and prepare_joint.py loaders accept these labels/frame directories through their explicit source/frames arguments. Native <dataset>/training.jsonl files preserve the recovery episode UID and provenance; RGB paths are relative to the dataset root (resolve before loading). Prefer these native manifests when mixing with original training data so the DAgger trajectory variant remains explicit. Do not treat these trajectories as independent oracle examples of the same episode ID.

Verification and provenance

SHA256SUMS.json lists every payload file. RELEASE_READY.json binds that inventory after complete local archive readback, label and image checks. audit/frames.jsonl.gz records PNG-byte and decoded-RGB hashes; audit/accepted_records.jsonl.gz retains executed/proposed/oracle actions, poses and committed action-history tokens. audit/merge_report.json records canonical coverage, exact source/model hashes and reviewed retry variants. audit/attempts.jsonl.gz lists every canonical ID and its selected outcome, with original/replacement record hashes. audit/canonical/ retains the exact source label/GT bytes; these are provenance inputs, not DAgger supervision. Some failed RxR attempts use the explicitly reviewed source-compatible body profile; successful originals keep their original profile. The R2R planner repair applies only to the reviewed failed retry. All use 15-degree turns. Local historical paths in the audit describe provenance, not download paths.

R2R metadata follows the preprocessed v1-3 labels linked by JanusVLN. RxR metadata follows DVLe/RxR_CE_15_deg_Trajectory_data. Underlying R2R, RxR and Matterport3D source terms continue to apply. Collection verification does not establish a trained-model performance gain.