RunLobster
Infrastructure for robot learning San Francisco

Infrastructure for
robot learning

Robot learning is held back by compute, not ideas. A $200 arm and an open checkpoint sit one 40 GB GPU away from working, and that gap is where most projects stop.

RunLobster builds the layer that closes it. The first piece is live:

The flagship from RunLobster Available now

Moving Atoms

Brains for Robots

Hosted state-of-the-art robot models you call over an API. Point an SO-101 at our endpoint and it takes plain-language instructions, with no GPU, no CUDA, no checkpoint downloads. Currently serving MolmoAct2, with 10M tokens free every day.

Serving MolmoAct2 10M tokens free / day No GPU, no CUDA SO-101 ready One API call
Try Moving Atoms → movingatoms.ai
Also in the works

The rest of the stack

Coming soon

Anvil

A brain of our own

A frontier robot policy, roughly 33B parameters, fine-tuned for real-world manipulation rather than benchmark leaderboards. Built to be served warm through the same endpoint, so it is one flag away when it lands.

~33B params · in development
Coming soon

Rehearsal

VR for Robots

A world model for policies to practise in. Today you find out a checkpoint is worse by breaking an actuator. We are building the imagined world that scores it first, so “is this better, and is it safe to run?” has an answer before anything moves.

World model · in development

Got an email from us?

We reach out to people working with LeRobot and SO-101 arms to invite them to try Moving Atoms, with unlimited hosted inference free for the trial period in exchange for telling us where it breaks. If that is why you are here, the product is real, it is free to try, and the person who emailed you is a human who will read your reply.

Not interested? Reply with “no thanks” and we will not contact you again, or write to hello@runlobster.com.