Figure's Helix 2.5 reaches 56% success in new homes

Figure's Helix 2.5 reached 56% success on household tasks in 30 homes it never trained on, per Figure's September 17 announcement. An identical policy trained from scratch on the same tasks managed just 9%. The 47-point gap comes from pretraining alone. The policy never collected data in the 30 homes and received no fine-tuning once it arrived.

Helix 2.5 runs on Figure's Index dataset, built from human behavior footage ingested at about 35 minutes per second. The same policy tidied living rooms, folded towels, and made beds without any per-home adaptation to the objects it manipulated. Figure says Helix 2.5 also needed 50% less task-specific data than its predecessor, Helix 02, while covering three times as many homes.

Figure doesn't say how it selected the 30 homes or the full list of tasks it tested. The company is also the only party that has evaluated Helix 2.5 so far, and Index's build involves a $3.5 billion compute commitment.