Skip to content

Troubleshooting·

Cameras·

/v2/front → 404 "Front frame not available"; /data fine. The SDK's headless Chromium is a participant in the rover's Agora channel. When the rover reboots, Agora fires user-unpublished, the page removes the #player-1000/1001 video elements and — in the upstream SDK — nothing recreates them. Our fork adds is_ready()/video_alive()/rejoin() and a watchdog loop (VIDEO_WATCHDOG=true, 5 s poll, 6 misses → fresh tokens + rejoin, backoff 30→300 s). If you run the upstream SDK, restart sdk is the manual equivalent.

Frames are black / tiny. ROVER_REJECT_BLACK_FRAMES, ROVER_BLACK_FRAME_STD, ROVER_MIN_FRAME_DIM govern what the recorder keeps; the SDK's IMAGE_QUALITY/IMAGE_FORMAT govern what it sends.

Driving·

rover_move returns but nothing happens. The firmware needs a continuous /control stream; the tools re-send frames for the whole duration. If a single frame is sent (e.g. from curl) the rover twitches and stops — that's expected.

Turns go the wrong way. ROVER_TURN_SIGN / DASH_TURN_SIGN flip the angular sign for the tools / the cockpit joystick.

The thinker never moves. Its drive flag is off (.memory/personas.json → thinker_drive, or SCOUT_THINKER_DRIVE). This is a feature — flip it in the cockpit's persona sheet.

Auth·

"Passkey not allowed" / ceremony fails. WebAuthn's relying-party id must be a hostname and the page must be HTTPS. Use scout.local (mDNS), an /etc/hosts name, or the tunnel hostname; set SCOUT_AUTH_RP_ID and SCOUT_AUTH_ORIGIN explicitly when behind a proxy.

Locked out. The store is auth.json inside the scout-auth-slim volume. Remove the volume and the first visitor re-enrols (set SCOUT_AUTH_BOOTSTRAP_TOKEN first if the rover is reachable from the internet).

401 on everything from a script. Mint a session with the auth module inside the dashboard container and send it as a Bearer:

TOKEN=$(docker exec scout-slim-dashboard python3 -c "import auth; print(auth.issue_token('cli','ops'))")
curl -sk -H "Authorization: Bearer $TOKEN" https://localhost:8080/api/telemetry

Voice·

401 invalid_api_key in the voice logs. The provider key was revoked (GitHub secret scanning does this to any key that touches a public place — or a "secret" gist). Rotate it in .env, --force-recreate voice.

Two things fighting over the mic. Only one voice persona; the listener and recorder must go through the MediaHub. Check scout-compose ps for a stray voice container and curl :8090/status.

Robot interrupts itself. Raise ROVER_GATE_TAIL_S (default 0.6 s) or tune ROVER_AEC_DELAY_MS.

Datasets·

Every episode replays the same video. lerobot 0.6 streams parquet and only writes the footer on finalize(). A recorder killed mid-session left the index footerless. Recent versions seal after every episode; for old datasets run the repair (tools/repair_episode_index.py --scan datasets, see Datasets).

"vcodec" TypeError / no dataset created. lerobot ≥ 0.6 moved video encoding to RGBEncoderConfig; _recorder_engine picks the right call per version (SCOUT_VCODEC, default h264). Make sure lerobot[dataset] (the extra) is installed.

torchcodec warnings. Harmless: it falls back to pyav.

Fleet·

No use_device in the agent. TINY_MCP=1 set? persona in TINY_MCP_PERSONAS? .tiny-mcp.env present and mounted? The bridge is fail-open: missing node/token/server → one warning, tools absent, persona still runs. The turn arrived from a sibling (/api/chat)? Then no fleet tools by design.

Containers·

Stopping the thinker truncated a parquet. Use the cockpit switch (checked between cycles) rather than docker rm -sf while an episode is open. Only the in-flight episode is lost since sealing.

docker compose says "no such service" for media/yolo. You forgot the override file. Use the scout-compose alias.

Ad-hoc commands in the image. The entrypoint dispatches on the service name, so for a shell use docker run --rm -it --entrypoint bash scout:slim or docker exec -it scout-slim-dashboard bash.