Inference
One command puts your robot policy on GPUs. Your robots call it from their control loop with connect() and act(); workers start as robots connect, stop when they go idle, and are billed by the second.
How it works
A deployment is a name your robots connect to. Behind it, the platform starts the workers, gives each robot a seat on one, and scales, replaces and versions them for you.
$ seq deploy pi05_droid:Pi05 --name my-pi05
import sequence_ai
with sequence_ai.connect(model="my-pi05") as policy:
out = policy.act("pick up the cup",
observe=robot.read, act=robot.apply,
max_seconds=30)connect() once, then act() directconnect() goes through the gateway once and returns the worker’s address. Every act() after it goes straight to the worker and brings back a chunk of future actions, the next one fetched while the last still plays, so the loop needs no round trip per step.
Workers start as robots connect, up to the GPU cap you set, and stop after ten minutes without a request unless you set another wait. They are billed by the second while they run; waiting for a cold one to start is not billed.
By default each robot gets a copy of the model of its own. Declare @seq.concurrent(max_concurrent=N) and one copy serves several robots: each worker seats as many as it can serve on time, and when every one is full another starts, up to your cap.
Start from an official template such as π0.5 DROID, GR00T N1.7 DROID or Cosmos 3 Edge DROID, or write one @seq.policy class. Every deploy is a new version that builds beside the one serving, takes over when it is ready, and can be rolled back at once.
Starter costs $0 a month and comes with $20 of free credit every month, renewed on the 1st. Usage is billed by the second, with no GPUs to reserve in advance.