Foundational infrastructure for machines that move
We run a policy distributed across cloud regions and GPU providers — close to every robot for low latency, resilient to any one outage — and benchmark it automatically on simulations you build for the environment it will run in.
Inference API
Every model behind one key — the ones that reason and the ones that move — with no GPUs to spin up. Scales to a managed tier for fleets in production: parallelised for throughput and priced down as you grow.
-
One envelope, both brains. Same
id,usage,perfandwarningswhether you asked a model to think or to move — only the payload differs. -
One
/actbody for every policy. Eight dimensions or seventy-five, one camera or three — the adapter absorbs it. - Parallelised for throughput. More policies batched onto each card as load rises, so the per-robot rate falls with scale instead of sitting at a fixed tier.
- Distributed across regions and providers. Capacity pooled across many GPU clouds and routed to the nearest region — latency stays low and no single provider’s outage stalls the fleet.
Simulation benchmarks
Benchmark a policy automatically on simulations you build for its target environment — every version scored on the conditions it will face, as you iterate. Priced per simulation GPU-hour, at the same GPU rates as inference.
- Your conditions, in simulation. Build the scenes the policy will meet — lighting, clutter, object positions — from your own environment.
- Automatic on every checkpoint. Each version runs the same suite as you iterate, so a regression shows up before it reaches an arm.
- One score to read. A pass rate per condition, with the failures grouped by what broke — not a wall of logs.
Call an action model the way you call an LLM
Every model is published the same way — what it reads, what it returns, and what it costs.
vlavla
vlavlawam
det
emb
depth
vlmvlm
vlmvlmvlm
vlm
vlm
decision