Researchers led by Changxuan Fan have built RoboTeam, a simulated benchmark in which AI agents control 4 to 16 robot arms in real time using only camera images, according to a paper on arXiv. Its 36 tasks are long jobs for several arms working as a team, for example handing timbers along or assembling a picture out of hundreds of cubes. The world does not pause while an agent thinks.
Using an adaptive system in which one agent decides when to hand work to helpers, gpt-6.1-sol completed 28 tasks and opus-5.5 completed 27, while four other models completed at most 3. For gpt-6.1-sol, a typical task lasted 3.5 hours and used 218 million input tokens, and a third of its steps had to be retried at least once.
Single-agent, hierarchical and decentralised systems completed a similar number of tasks. The hierarchical and decentralised systems were faster but used up to 1.9 times the tokens, and the single agent used fewer tokens but took 1.6 times as long. No model or system was yet fast, cheap and reliable at once, the authors write.



