RoboTeam benchmark tests AI agents controlling up to 16 robot arms in real time
A new simulated benchmark from arXiv asks multimodal agents to coordinate many robot arms on long shared jobs, using camera images only, with no pausing and no ground-truth positions.
In plain words
Researchers introduced RoboTeam, a simulated testbed where AI agents direct between 4 and 16 robot arms at once. The agents see the world only through camera images and must finish long jobs that require the arms to cooperate. The benchmark highlights how slow, costly and error-prone current models remain at this kind of control.
What was announced
The authors say existing benchmarks fall short in three ways: they usually test one or two arms on short tasks, pause the simulated world while the agent thinks, and hand the agent ground-truth object positions. RoboTeam instead offers 36 long tasks that arms must complete together, such as passing timbers or building a picture from hundreds of cubes, while the world runs in real time and perception comes from camera images alone.
The benchmark starts from an adaptive agent system in which one agent decides when to fork delegate agents. With that setup, the models gpt-6.1-sol and opus-5.5 completed 28 and 27 of the 36 tasks, while four other models managed at most 3. The cost is high: a typical task took gpt-6.1-sol 3.5 hours and 218 million input tokens, and a third of its steps failed at least once and needed retries. Compared with fixed single-agent, hierarchical and decentralized systems, all four completed a similar number of tasks but traded off time and tokens; hierarchical and decentralized setups were faster but used up to 1.9 times the tokens, while the single agent used fewer tokens but took 1.6 times as long. The authors conclude that no model or system is yet fast, cheap and reliable at once.
What it does not show
- The evidence is the paper's abstract only; full experimental details, methods and results are not available in the supplied passages.
- Results are from a simulated benchmark and reflect the tested models and agent systems; no independent confirmation is provided.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.