Key points

  • RoboBuilder has AI coding agents design a robot's hardware and program from standard parts, motors and one custom 3D-printed piece.
  • Researchers tested four frontier coding agents and built 135 physical robots in an estimated 240 hours.
  • Across all difficulty levels, 66% of designs that passed in simulation also worked on the real course.
  • After watching video of a real failure, agents redesigned their robots and eliminated 72% of level-1 and level-2 failures.

AI agents built to write software can also design working robots, but many of those robots fail once they leave the computer, according to a benchmark called RoboBuilder. The work was posted to the arXiv preprint server and has not yet been peer-reviewed.

In RoboBuilder, an AI agent designs both a robot's body and its control program, using off-the-shelf parts, motors and one custom 3D-printed piece. Each design that succeeds in simulation is then assembled and run on a physical obstacle course. The team tested four frontier coding agents on ten obstacles across three difficulty levels. That meant running 16,427 simulations covering 12,774 distinct robots and building 135 physical machines in an estimated 240 hours.

The agents cleared nearly all of the easy obstacles and came up with mechanisms that the team's automated search methods never found. They struggled on harder levels and when a single robot had to get past every obstacle. They also struggled in the real world: across all levels, 66% of designs that passed in simulation also worked on the physical course. Early on, using publicly available part data, none of the first 15 designs that passed in simulation crossed the real course. So the researchers let agents request measurements and watch videos of real runs to improve the simulator.

That feedback helped. After seeing video of one of its robots failing, an agent would redesign the robot, sometimes correcting its simulator as well. This removed 72% of failures at levels 1 and 2. The authors present the benchmark as a way to track progress toward AI agents that can engineer real machines.

Why it matters

The researchers present RoboBuilder as a way to measure progress toward AI agents that can engineer real machines, not just designs that work in simulation.

What we don't know yet

The study is a preprint and has not been peer-reviewed. The agents struggled at harder difficulty levels, when one robot had to cross every obstacle, and in moving designs from simulation to the real world.

Terms in this story

coding agent
An AI system that writes and runs computer code on its own to complete a task.
co-design
Designing a robot's physical body and its control software together rather than one after the other.
benchmark
A standard set of tasks used to measure and compare how well different systems perform.

Sources and verification

  1. RoboBuilder: Agentic Robot Co-Design with Real-world Video Feedback in the Loop arXiv · Nov 19, 2026 · preprint, not peer-reviewed
Fact-checked against the source on November 20, 2026: 20 factual claims verified.
Written in our own words: longest phrase shared with the source is 11 words.
Illustration created by DexScience; no third-party image was used.

DexScience articles are produced with AI assistance and an automated fact-check against the linked sources. Spotted an error? Tell us and we will correct it openly.