← Latest papers
💻 computer science

ROSA: A Robotics Foundation Model Serving System for Robot Factories

ROSA is a robotics foundation model serving system designed for robot factories that leverages shared GPU pools, multi-model programming abstractions, and factory-objective-driven scheduling to maximize overall productivity, achieving up to a 12.06x improvement over conventional dedicated serving systems.

Original authors: Wenqi Jiang, Jason Clemons, Rowland O'Flaherty, Hugo Hadfield, Alperen Degirmenci, Shuran Song, Yashraj Narang, Christos Kozyrakis

Published 2026-07-02
📖 4 min read☕ Coffee break read

Original authors: Wenqi Jiang, Jason Clemons, Rowland O'Flaherty, Hugo Hadfield, Alperen Degirmenci, Shuran Song, Yashraj Narang, Christos Kozyrakis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a factory floor filled with dozens of robots, each trying to do a specific job like picking up parts, assembling boxes, or checking for defects. In the past, the standard way to run these robots was to give every single robot its own personal brain (a powerful computer chip) sitting right on its back.

The paper argues that this approach is like giving every employee in a massive office their own personal supercomputer. It's expensive, wastes energy, and often leaves those supercomputers sitting idle while the employee is just walking to the printer.

Instead, the authors propose ROSA (Robotics Oriented Serving Architecture). Think of ROSA not as a personal computer for each robot, but as a centralized "Brain Cloud" that all the robots share.

Here is how ROSA works, broken down into three simple ideas:

1. The Shared Brain Pool (Instead of Personal Brains)

The Old Way: Each robot has its own small, weak computer. If the robot needs to do a complex calculation, it struggles because its battery is small and the computer isn't very fast.
The ROSA Way: The robots send their "questions" (like "What should I grab next?") over the internet to a central room full of massive, super-powerful servers.

  • The Analogy: Imagine a restaurant. In the old way, every table has its own tiny kitchen with a single chef. In the ROSA way, there is one giant, high-tech kitchen in the back. The waiters (robots) just send orders to the kitchen. The kitchen can cook many orders at once, uses better equipment, and saves energy because the waiters don't need to carry heavy stoves around.
  • The Result: The robots can run longer on their batteries (since they aren't powering a heavy computer) and get answers much faster because the central kitchen is much stronger than any single robot's brain.

2. The Smart Manager (The Scheduler)

Just having a big kitchen isn't enough; you need a smart manager to decide who cooks what and when.

  • The Problem: A robot factory isn't just about moving fast; it's about not making mistakes. A robot might need a "Safety Checker" (to make sure it doesn't hit a human), a "Planner" (to figure out the steps), and an "Action Taker" (to actually move).
  • The ROSA Manager: ROSA acts like a traffic controller. It looks at all the robots and says:
    • "Robot A needs a safety check right now, so let's prioritize that."
    • "Robot B is doing a simple task, so we can wait a second before sending its next order."
    • "Let's group the orders from 10 robots together and send them as one big batch to the server."
  • The Goal: The manager doesn't just try to make one robot fast. It tries to make the whole factory productive. It ensures that the most important tasks get done and that no robot crashes because it waited too long for an answer.

3. The "Safety Net" Rules

The paper explains that ROSA has built-in rules for when things go wrong, similar to how a human supervisor would react.

  • If a robot is about to hit a human: The safety model says "Stop!" and the robot immediately freezes or moves to a safe spot.
  • If the server is too slow: If the central brain takes too long to answer, the robot doesn't just wait forever; it has a backup plan (like stopping or asking a human for help).
  • If a task fails: If a robot drops a box, the system notices, resets the robot, and tries again. If it keeps failing, it calls a human worker.

What Did They Find?

The researchers tested ROSA on real robots (a Franka Panda arm) and simulated large factories with up to 64 robots.

  • The Big Win: Compared to the old method where every robot had its own dedicated computer, ROSA made the factory 12 times more productive.
  • The Comparison: Even compared to other systems that shared a server but didn't have the "Smart Manager," ROSA was 2.4 times more productive.

Summary

ROSA is a new way to run robot factories. Instead of giving every robot a heavy, expensive, and underused computer, it connects them to a shared, powerful cloud of computers. A smart scheduler manages the traffic to ensure the whole factory runs smoothly, safely, and efficiently, getting the most work done with the least amount of wasted energy.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →