DreamZero DROID troubleshooting

Start by separating session setup, camera delivery, model output, and robot execution. A correctly shaped chunk establishes a transport result; it does not establish useful or safe motion.

Symptom	Check
HTTP 401/403 or model inaccessible	Confirm `REACTOR_API_KEY`, endpoint, and model access; avoid sharing the key in logs.
HTTP 429 with `no available capacity`	No capacity is available. Retry with bounded backoff or contact Reactor; avoid rapid reconnect loops.
Slow `READY`	Cold startup can take minutes. The quickstart allows 900 seconds; retain status transitions and the session ID when escalating.
Ready but no `episode_started`	Publish `exterior_1`, `exterior_2`, and `wrist` after `READY`, deliver at least one frame per view, and send a nonempty prompt. Inspect `command_error`.
One chunk, then silence	Every camera must provide a new frame before another chunk can run. Check the least active track.
Disconnect while waiting	The SDK manages keepalive; do not block the event loop with synchronous control work.
Joint targets jump after reset	Reset clears measured state. Republish actual joint and gripper positions before starting the next prompt.
A chunk seems one observation behind	An older inference can already be in flight. `obs_seq` is a server ingest counter, not your capture ID.
Wrong shape or nonfinite values	Reject the chunk locally. Expect `(24, 8)`; record release, counters, and shape for support.
Black `predicted_video`	Preview is disabled by default. Enable `set_video_preview` only if you need diagnostic images.
Primary camera appears black	Check the RoboLab index offset: `exterior_image_0_left` maps to `exterior_1`. The default black second view belongs on `exterior_2`.
First simulation request stalls	Keep lazy gateway connection enabled and check camera track startup. Run `check_wiring.py` before diagnosing model quality.

Report a reproducible issue

Include the model identifier, deployed release, SDK version, session ID, status transitions, camera names and shapes, command sequence, chunk_index, obs_seq, and timing boundaries. Include whether the input was synthetic, recorded, simulated, or live. Do not attach API keys or private camera footage unless explicitly approved for sharing. Measure first-chunk startup separately from steady-state inference and client transport. The inference_seconds field is not total observation-to-action latency. If hardware is connected, use the controller’s hold/stop behavior while investigating; changing or resetting the model does not stop the robot.

Get started

Models

DreamZero DROID troubleshooting

Report a reproducible issue

​Report a reproducible issue

Report a reproducible issue