A robot assembly system achieves 82% success on sub-millimeter tasks.

Robots struggle to assemble tiny parts because they cannot properly feel or predict physical contact. Researchers built Facet-0, a foundation model that helps robots handle these delicate, sub-millimeter tasks. It works by combining vision, language, and movement history to predict both its next action and the forces, or wrenches, that action will create. It uses reinforcement learning, where the system learns through trial and error, guided by a critic that evaluates contact outcomes. A specialized lightweight actor then adapts the frozen model directly on the robot. Tested on five computer-assembly tasks, Facet-0 achieved an 82% mean success rate, while the strongest baseline only reached 15%. It placed parts with 0.5 mm accuracy at a 50 ms command latency.

Why it matters

This matters for manufacturers who need robots to assemble tiny, precise parts like in computers. With 82% success, it makes robotic assembly far more reliable, reducing errors and improving production efficiency.

Questions answered