AI agents can operate a robot arm, but simple precision still defeats them
In an eight-task robotics preview, all five tested systems pressed a button successfully, while none fully completed the box-measurement task.
Artificial Analysis’s October 14 robotics preview gave five coding-agent systems control of a robot arm and eight tabletop tasks.
All pressed a button and placed a toy correctly; none achieved full marks for measuring a box within the specified tolerance. [1]
GPT-6 Astra led with 69.4 out of 100 on the aggregate score, which includes partial credit. That is not a 69.4% task-success rate. It averaged about 30 minutes per task, against a human average of about 20 seconds in this test. [1]
The useful signal is the gap between manipulating objects and doing so precisely and efficiently. This small preview, using different native agent software, does not establish industrial reliability. The published task recordings are a better guide to failure modes than the ranking alone. [1]
Sources & details 1 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Artificial Analysis · AA-Robotics v0.1
Original experiment and task recordings. Five systems, eight tasks, high-effort native agent configurations and partial-credit scoring; no independent replication inspected.
Images & figures
AI-generated editorial illustration of the difference between gross manipulation and exact measurement; not documentary evidence or a measured diagram.
Euangelion · AI-generated editorial illustration · Source · Usage terms