navila-teleop-odom

Teleoperated NaVILA-format episodes collected on a Unitree Go1 + ZED stereo

• Jetson AGX Orin stack, for LoRA fine-tuning the NaVILA-Llama3-8b-8f checkpoint with an ask-user clarification capability.

Collection pipeline, schema details, and the broader research context live in the Conversational_VLA_Navigation repo (CLAUDE.md).

Sessions

Session	Episodes	Samples	Notes
lab_test/	45	235	Original lab collection
lab_test_mirrored/	45	235	Horizontal flip of lab_test
corridor_test/	18	203	Original corridor collection
corridor_test_mirrored/	18	203	Horizontal flip of corridor_test
Total	126	876

Mirrored sessions are produced by data_collection/mirror_dataset.py (left↔right swap on JPEGs, action labels, instruction text, telemetry yaw rate). Doubles the L/R class distribution; same locations and trajectories as originals.

Format

Per-episode files (the canonical format on the collection device):

<session>/<episode>/
├── annotations.json    # list of training samples (q, a, frames, clarification, ...)
├── meta.json           # per-episode bookkeeping (instruction transitions,
│                       # ask_user_events, amender_events, location, scene, ...)
└── frame_NNNN.jpg      # ZED frames (only frames referenced by annotations
                        # are pushed to this repo; raw collection has more)

Each sample in annotations.json:

{
  "video_id": "episode_NNNN_YYYYMMDD_HHMMSS",
  "frames":   ["frame_0000.jpg", ...],   // 8 entries, NaVILA memory-bank layout
  "q":        "<navigation instruction>",
  "a":        "The next action is move forward 50 cm.",
  "a_bare":   "move forward 50 cm",      // pre-migration label, audit-only
  // ask-user samples additionally carry:
  "clarification":            "...",
  "clarification_consistent": "...",
  "clarification_varied":     "...",
  "clarification_source":     "vlm" | "operator",
  "operator_question_raw":    null | "...",
  "oracle_instruction":       "...",
  "oracle_used_pruned":       false,
  "clarification_answer":     null | "<operator's reply>",
  "amended_instruction":      null | "<rewritten instruction>",
  "telemetry":                {"measured": {"vx_mps": 0.0, "wz_rps": 0.0}}
}

Action label space (always written in the prefixed form to match the base NaVILA checkpoint's output distribution):

• The next action is move forward N cm.
• The next action is move backward N cm.
• The next action is turn left N degrees.
• The next action is turn right N degrees.
• The next action is stop.
• The next action is ask user.

Use this dataset for training

NaVILA's LazyVLNCEDataset consumes a flat <session>/annotations.json that concatenates every episode's per-sample records and rewrites the frame paths to be session-relative (<episode>/frame_NNNN.jpg). The flat files are not included in the push — generate them after pulling, then register each session with NaVILA's data mixture, in one command:

# After downloading the dataset, from the dataset root:
./register_sessions.sh --navila-repo /path/to/NaVILA

This bundled helper:

1. Runs build_flat_annotations.py for each of the four sessions — reads every <episode>/annotations.json, defensively filters out malformed samples (empty q, non-final a, wrong window size, missing referenced JPEGs — drops are logged), writes <session>/annotations.json ready for LazyVLNCEDataset.
2. Calls <NaVILA>/data_collection/navila_odom_collector.py --register for each session, registering them under distinct names (navila_teleop_odom_lab_test, ..._lab_test_mirrored, ..._corridor_test, ..._corridor_test_mirrored) in NaVILA's llava/data/datasets_mixture.py.
3. Prints the training-ready --data_mixture string at the end.

If you only want the flat files (e.g. for a one-off eval, or your training harness doesn't go through NaVILA's data mixture), pass --skip-register:

./register_sessions.sh --skip-register

build_flat_annotations.py is stdlib-only and idempotent — safe to re-run after pulling incremental updates.

Pull commands

# Read-only pull (set HF_TOKEN in env if the repo is still private)
huggingface-cli download SunnyDeshpande/navila-teleop-odom \
    --repo-type dataset \
    --local-dir <local_clone> \
    --local-dir-use-symlinks False

--local-dir-use-symlinks False materializes JPEGs into the local dir; needed if the cluster's compute nodes don't share the HF cache mount.