# RoboBuilder tests AI-designed robots on a physical course

> The benchmark builds robots that pass simulation and tests them in the real world. Video feedback helped agents fix many failures, but harder tasks remained difficult.

Oossa · 2026-11-19 · https://oossa.com/en/robobuilder-tests-ai-designed-robots-on-a-physical-course

Researchers introduced RoboBuilder, a benchmark where coding agents design both a robot’s hardware and its control program using standard parts and a custom 3D-printed piece. Designs that pass simulation are built and tested on a physical obstacle course.

In the researchers’ tests, none of 15 initial designs that passed simulation completed the real course. Across four coding agents, 135 robots were built; simulated successes carried over to real trials 66% of the time across all difficulty levels.

When shown videos of real-world failures, agents redesigned robots and sometimes corrected their simulator, removing 72% of failures on the two easier levels. They still struggled with harder obstacles and with designing one robot that could cross every obstacle.

## The facts

- The study reports 16,427 simulations covering 12,774 distinct robot designs.
- The paper was published on arXiv on 2026-11-19.

## Why it matters

For robotics researchers using simulation to screen designs, RoboBuilder shows that a passing simulation is not enough: in these tests, about one in three simulated passes failed to hold up across the real-world trials.

## Sources & references

1. [RoboBuilder: Agentic Robot Co-Design with Real-world Video Feedback in the Loop](https://arxiv.org/abs/2611.23855) – arXiv, 2026-11-19

Last updated: 2026-11-19
