Web Manager

RobotDisco VLA Manager (webui/)

A browser console for the whole loop between recording and deployment, with three tabs — Datasets (browse/merge/delete/import, per-episode trim/delete/rate/annotate/rename-task/move, and an episode player that renders each episode to a static rerun .rrd shown in rerun's WASM viewer over plain HTTP), Models (every training run + live progress for the one training now), and Evals (each eval run with rating/notes + metrics). The ops delegate to lerobot.datasets.dataset_tools and lerobot.scripts.lerobot_dataset_viz (no bespoke dataset surgery except sub-episode trim, which has no library equivalent). It's self-contained (only needs lerobot + FastAPI), so it runs the same on the GPU box, your Mac, or in Docker.

Run it

Native (GPU box or Mac) — one-time uv pip install -r webui/requirements.txt into the project venv, then:

HF_HUB_OFFLINE=1 PYTHONPATH=. .venv/bin/python -m webui.server \
  --host 0.0.0.0 --port 8080 --datasets-dir outputs/datasets
# GPU box: browse from the Mac at http://192.168.11.130:8080  ·  Mac: http://localhost:8080

Docker (multi-arch amd64/arm64; outputs/datasets is mounted, not baked in, so edits persist on the host):

docker compose -f webui/docker-compose.yml up --build       # then http://localhost:8080
# portable image for the x86 GPU/robot infra AND the Mac:
docker buildx build --platform linux/amd64,linux/arm64 -f webui/Dockerfile -t <registry>/vla-webui:0.4.4 --push .

The image bundles lerobot + ffmpeg (CPU-only torch). On a Linux host uncomment user: "${UID}:${GID}" in the compose service so volume files aren't root-owned; the import-from-robot-PC button needs your SSH key mounted (~/.ssh:/root/.ssh:ro).

Tabs

• Explorer — cards per dataset (episodes/frames/fps/size/task/media); check ≥2 → Merge; + Add imports a dataset from the robot PC (rsync from rd@.80) or copies a local path; the header polls the folder every 4 s and flags newly-recorded/-changed datasets.
• Episode manager — per row: ▶ (play), trim (drop a time window), move (to another dataset), del, click-to-set rating stars, inline notes + operator. Rename-task applies to the whole dataset or just the selected episodes.
• Player — ▶ renders the episode to a rerun .rrd (first play of an episode ~30s, then cached) and loads it in rerun's WASM viewer over ordinary HTTP — no gRPC, so it works in any browser. Needs ports 8080 (UI + the .rrd) and 9090 (viewer app) reachable; the Docker setup publishes both.
• Models tab — lists every run under outputs/train/ (policy, dataset, steps, checkpoints, size) and shows the actively-training one live (step/loss/ETA, auto-refreshing); finished models get download (zipped checkpoint) / rename / delete.
• Evals tab — each eval run as a card with ★ rating, notes, operator + metrics (model, checkpoint, task, episodes, success rate). Bridge evals are auto-captured GPU-side into outputs/evals/ by scripts/infer_server.py; you can also add manual entries. Rename / delete / download / play-video per eval.

Safety

Rename-task + annotations are in-place; delete/trim/merge/move rebuild the dataset and keep the pre-edit copy as <name>.bak (one-click restore .bak). Destructive calls require a confirm. move pre-checks fps/feature compatibility before touching either side, so a rejected move can't half-apply. Ratings/notes/operator live in a sidecar (meta/annotations.json), outside the LeRobot schema, and are remapped through the new episode indices on every rebuild.