The upstream model card describes
the DROID policy and its checkpoint variants. Compatibility requires the DROID camera, state, and
action conventions, not only a seven-joint arm.
Pricing
$0.102 per minute of session time. Prices are in USD, metered per second from when the session becomes ready until it terminates. Time between action requests is billable. See Pricing & Billing for details.Checkpoint choices
Reactor measured these medians on November 4, 2026, using release
0.3.0 and Python SDK 1.13.0.
Each checkpoint used a fresh session and 20 predictions after excluding its first prediction, with
identical synthetic observations, a fixed task and seed, and three 640 × 360 camera streams at 15
fps. Values come from the reply’s inference_seconds field; they exclude client transport and robot
execution and are not latency guarantees.
Use these measurements to shortlist checkpoints for your latency budget, then evaluate task success
on your own robot or simulator to choose the latency–quality tradeoff that fits your application.
This benchmark measured inference latency, not policy quality. Faster inference does not always
produce a faster client response; measure both in your integration.
For a first run, use the available default returned by get_checkpoint, as the quickstart does. To
select another checkpoint, call get_checkpoint and inspect checkpoint_selected.available before
selecting. The API defines six choices, but your session may have only some of them available.
select_checkpoint locks your choice for the session. Selecting the same checkpoint again is
accepted; choosing a different one requires a new session, even after reset. If you do not select
one, the first valid request locks the default. Check the reported checkpoint to confirm your
choice.
Request your first actions, then implement the
simulator bridge. Use the
API reference for exact fields and payloads.