MEASUREMENT BOUNDARY
One wall-clock value cannot describe the whole embodied loop.
Record monotonic timestamps for observation capture, availability, preprocessing, inference start/end, action decoding, dispatch and observed feedback. Compare them without confusing dispatch with actuator onset.
On one computer, monotonic nanoseconds make interval measurements interpretable. Across robot client and remote inference server, record an independent clock alignment and UTC anchor; never subtract unrelated machine-local monotonic clocks as if they shared an origin.
METRICS
Publish distributions and data age, not just an average FPS.
The experimental schema reserves capture-to-dispatch latency, model compute latency, dispatch-to-feedback latency, p50/p95 of important intervals, stale frame age at the action boundary, queue-empty events and dropped frames.
An action stream running at a target rate does not prove every joint moves at that rate. Likewise, reported camera FPS does not establish that each frame reaches the policy with a known bounded delay.
Timing distribution
Characterize normal response and tail latency rather than hiding stalls in a mean.
Staleness
Measure how old the observation was when its corresponding action was dispatched.
Action continuity
Count exhausted action queues and overlapping chunks; distinguish continuity from timely corrections.
Physical observation
Record return measurements separately from issued targets and software command times.
SYNC VS ASYNC
Asynchronous inference changes idle time and the meaning of freshness.
Official LeRobot async inference separates a PolicyServer from RobotClient. The client can execute one action chunk while the server generates the next. Relevant parameters include actions_per_chunk, chunk_size_threshold and action aggregation.
The tradeoff is not simply “async is faster.” Old actions may remain in a queue after the world changes, and large chunks can reduce reactiveness. A valid SO-101 test must evaluate queue starvation, action age and task correction under matched perturbations.
OPENRGD
The semantic/runtime boundary must survive slow or remote inference.
A high-level model should not become final actuator authority merely because it emits continuous numbers. In OpenRGD’s candidate architecture, proposed actions pass through embodiment grounding and execution-responsibility boundaries outside the canonical standard repo.
No OpenRGD SO-101 latency numbers, remote 5080 inference speed or 1 kHz hardware-control performance are claimed here. Those need timestamps from the actual implementation.