Diagnose missing action chunks, step acknowledgements, and observation mapping.
Symptom
Check
Connection rejected or no capacity
Use https://api.reactor.inc, the correct model identifier, and an authorized key. Capacity errors can occur before inference; retry with backoff or contact Reactor.
No first chunk
Wait for READY before publishing. All three named tracks must deliver frames, and proprio_json must parse as row lists.
No next chunk
Acknowledge the last received step after execution. Repeating or lowering the acknowledged counter does not open the gate.
State ignored
Encode JSON as a string inside the named command payload. Check the 8,000-character limits and row widths.
Correct shape, wrong motion
Check camera identities, joint ordering/units, and absolute-target interpretation. Tensor shape does not prove embodiment compatibility.
Gripper moves the wrong way
DROID uses 0 open and 1 closed. Convert your driver’s units/polarity explicitly.
Values differ on identical inputs
Sampling is unseeded and video transport can change the received pixels; exact numerical equality is not a supported check.
Unexpected step after reset
Reset restarts the counter but preserves input fields. Use a new session when old acknowledgements or replies could be mistaken for new ones.
Keep the event loop responsive so camera publishing and message handling can continue. Start a fresh
session after an uncertain timeout; do not fabricate an acknowledgement to force progress. The
documented endpoint supports reset as specified in the
reference.
Include the model identifier, SDK/client versions, session ID, timestamps, track names/shapes,
proprioception dimensions, last received and acknowledged counters, and exact error text. Never
include an API key. State whether the failure occurred during connection, first prediction, or
subsequent control steps.
⌘I
Assistant
Responses are generated using AI and may contain mistakes.