A paper posted to arXiv on Sept. 30 introduces AssemblyWorld, an interactive 3D test of whether general-purpose AI agents can assemble objects through visual inspection and tool use. The top system completed 59.4% of whole-object tasks, even though it placed 80.9% of individual parts correctly.
The scores measure different levels of success. Part accuracy counts how many parts land within the study’s geometric tolerance; complete-assembly success requires every part of an object to pass. A mostly correct arrangement can therefore fail the full-object test because of one remaining placement error.
Agents worked in a simulator, not with physical hardware. They could request rendered 2D camera views and use tools to move or rotate rigid parts. The system exposed object IDs and pose and size records, but hid the underlying mesh vertices and faces. Some tasks included a reference image or assembly manual; broken-object reassembly tasks had no picture of the intact object.
AssemblyWorldBench contains 100 tasks across 80 objects, covering furniture assembly, industrial assembly and reassembly of broken objects. The researchers tested eight agent systems with the same tools, starting arrangements, instructions and 60-minute limit. GPT-6 Astra had the highest whole-object success rate. Reference images helped some stronger systems, but the size of that benefit varied by system and task.
The benchmark measures final geometry in a controlled 3D environment; it did not test collisions or whether an object would remain physically stable. The study did not test whether agents can carry this accuracy into physical assembly. Its sampled tasks and unknown overlap with training data also limit broad claims about generalization.
Sources
- AssemblyWorld paper, full text
- AssemblyWorld project page
- Illustrative photo by Stilfehler, showing IKEA Sektion cabinet-frame components during assembly. Wikimedia Commons file; licensed under CC BY-SA 4.0. The image is not from the study.
