Figure AI

Robotics · AI · Humanoid Robots

Milestone

Figure's Helix 2.5 humanoid hits 56% success in 30 unseen homes

Household robots typically need retraining per home; here, human video pretraining, not robot demos, made a single policy generalize instantly.

• Figure unveiled Helix 2.5, a neural network pretrained on its Index human-behavior dataset, and had it complete three household tasks zero-shot across 30 unseen Bay Area homes with no data collection or fine-tuning on-site.
• Figure builds humanoid robots like Figure 03 for logistics and home settings; its prior Helix 02 model needed data collected at each deployment site, including a 200-hour autonomous dishwasher-unloading run.
• In a controlled comparison holding architecture and data fixed, Index pretraining lifted zero-shot task success from 9% to 56% (237 of 420 trials) versus a policy trained from scratch.
• Each of the three tasks (tidying, towel folding, bed making) used one fixed checkpoint across all 30 homes, required full task completion for credit, and counted any safety intervention as a failure.
• Helix 2.5 needed only half the task-specific training data of a comparable Helix 02 behavior yet generalized to 30 unfamiliar homes, roughly a 30x expansion in deployment scope.
• Figure has committed $3.5 billion in compute to the Helix program, while its Index dataset now adds about 35 minutes of human-behavior video per second and has paid contributors $15 million.
• The result suggests pretraining humanoids on human video, rather than site-specific robot data, could let household robots scale to new homes without costly per-site data collection, though a 56% full-task success rate still leaves a real reliability gap before dependable service.