Figure says its Helix 2.5 humanoid policy completed three predetermined household tasks in unfamiliar homes 56% of the time, compared with 9% for a matched policy trained from scratch. The result is meaningful evidence of environment transfer, not an open-ended household robot.\n\n### Key facts\n\n- Figure reports 56% all-or-nothing success across 30 unseen Bay Area homes, versus 9% from scratch.\n- Tasks were living-room tidying, towel folding and bed making.\n- One fixed checkpoint per task ran without home-specific adaptation.\n- Primary source: Figure's Helix 2.5 report.\n\nZero-shot needs qualification. Figure says Helix 2.5 was pretrained on Index, then adapted with task-specification data for these three behaviors. The homes and specified objects were held out. This tests a learned policy's ability to transfer physical and visual priors into a new setting, not its ability to learn arbitrary chores on arrival.\n\nThe rubric is strict. All scattered toys must reach the basket; every towel must be folded and placed; bed making requires both pillows and both comforter corners positioned correctly. There is no partial credit. This makes 56% more informative than a successful demo reel, but it also means nearly half the end-to-end trials failed.\n\nThe report spans locomotion, active perception, bimanual work and deformable objects. The 56-to-9 comparison is the key evidence that broad behavior pretraining helps transfer. Several limits remain: Figure does not say how homes were selected, how task data was collected, or whether social-video footage constitutes a full record of trials.\n\nFigure calls the result zero-shot generalization, which is accurate within the evaluation. The wrong translation is that a robot can now handle housework generally. The strongest counterargument is reliability and sampling; 56% is research progress, not service quality. The next credible test is independent replay across diverse homes, with disclosed data provenance and recovery behavior.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)