Figure AI published an evaluation of its Helix 2.5 model on 17 Sep 2026. The robot worked in 30 Bay Area homes, and, according to Figure, “No data was collected in any evaluation home.”
The test covered three household tasks: tidying living rooms, folding towels and making beds. Figure counted a trial as a success only when the whole task was done: every item tidied, every towel folded or the whole bed made. The policy pretrained on Figure’s Index dataset succeeded in 56% of zero-shot trials; a policy trained from scratch succeeded in 9%. The post also says Figure has committed $3.5 billion of compute to training Helix.
The evaluation was run and reported by Figure; no outside party has published a replication. The post does not name the robot model used. Robopear records the result as a vendor demonstration; no Figure card changes status with it.