Robots finish about half of everyday chores in makers' own tests, and 12% in an independent one
Figure and Physical Intelligence report about half of chores fully done in their own tests. In an independent challenge, the best got 12%.
Three recent tests show where general-purpose robots stand on everyday chores. In September 2026 Figure tested its Helix 2.5 model in 30 Bay Area homes where it had collected no data. The tasks were tidying a living room, folding towels and making beds. Its humanoids completed the whole task 56% of the time, with no credit for partial work. The same model trained without Figure’s large pretraining dataset (the broad data a model learns from before any specific task) managed 9%.
Physical Intelligence put a fine-tuned π0.6 model through the “Robot Olympics”, a public list of hard everyday tasks such as turning a sock inside out, using a key and cleaning a greasy pan. Average success was 52% (December 2025). A standard AI model fine-tuned on the same data, but without large-scale robot pretraining, succeeded on none.
Independent tests are fewer. In Stanford’s 2025 BEHAVIOR Challenge, 18 teams controlled a simulated robot through 50 full household tasks such as making toast or tidying a room. The winning team fully completed 12.4% of the test tasks (26% counting partial progress). The 2026 challenge doubles the list to 100 tasks; we had not found its results when this entry was added.
What it means
In the makers’ own comparisons, large-scale training is what lifts these robots from almost never finishing a chore to finishing about half. But a robot that fails half the time is not ready to be left alone in a home, and the independent test scores lower still. For comparison, Figure’s own target on a car factory line was over 99% success.
Caveats
- Figure and Physical Intelligence chose and ran their own tests, and neither result has been reproduced by anyone else.
- About half is far below what paying customers expect. At BMW, Figure’s own target for loading car parts was over 99% placement success per shift.
- Physical Intelligence says its results varied between runs and that it did not tune for the highest success rate; it fine-tuned with usually under nine hours of data per task.
- BEHAVIOR runs in simulation, so its scores do not transfer directly to real robots, and its tasks differ from the companies’ tests.