An AI practiced, took notes, then tackled 14 household tasks a robot policy struggled with

Instead of updating its model weights, an AI practiced household chores in a simulator and wrote instructions for its next attempt. In a paper submitted November 10, Robot Learning Collective reports that GPT-6 Astra reached a mean goal-completion score of 0.639 on 14 selected difficult tasks, compared with 0.108 for the team's trained robot policy. The score measures the fraction of goal conditions satisfied, not the percentage of entire chores completed.

Practice broke each task into troublesome steps such as grasping, pressing and opening doors. After repeated failures, the agent examined human demonstrations, compared the motions and tried again. Working commands and measurements became written procedures. Evaluation started from a frozen copy of those records. The agent could inspect camera images, depth and robot state, then write short programs calling motion tools. Practice-only privileges, including restoring saved scenes and checking goal conditions, were removed.

The system entered the 2026 BEHAVIOR Challenge, which covers 100 simulated household activities from switching on a radio to preparing food. A learned action model still handled most tasks; the team selected 14 for the coding agent after its own trials. It reports ninth place overall and a 27.3% full-task success rate for the final submission. The interesting result is the connection between a general-purpose AI's trial-and-error reasoning, written experience and robot control.

That connection is still slow and costly. Agent runs took a median of about 38 minutes of wall time, roughly 17 times the simulated duration, and cost tens of dollars in API fees. The simulator paused while the AI thought. Pouring, precise placement and repetitive loops remained problems. These are team-reported results using the official evaluator, without organizer retesting on the private test set, and the 14 tasks were not randomly selected. A practical home robot remains far away. A nearer research question is whether the useful lessons from expensive practice can be turned into fast, affordable actions.

Source time: arXiv initial submission, November 11, 2026, 02:06:16 Beijing time / November 10, 2026, 13:06:16 EST.