NewsSep 17, 2026, 20:21 UTC

Figure Helix 2.5: zero-shot work in 30 unseen Bay Area homes

Key takeaways from Figure's Helix 2.5 write-up and founder Brett Adcock's 17 September 2026 thread on Figure:

Zero-shot in 30 unseen homes: Figure rented 30 Bay Area homes and ran Helix 2.5 with no data collection, fine-tuning or adaptation in those environments or on the objects being manipulated. Adcock posted four hours of the robot doing that work. The three tasks were tidying living rooms (13–15 toys into a basket), folding towels into a basket, and making beds (pillows and comforter corners to the top of the bed, comforter pulled smooth).
One foundation model, three whole-body behaviours: A single Index-pretrained base model was adapted to locomotion, rigid and deformable manipulation, bimanual coordination and active perception. Helix 02 had shown long-horizon whole-body control, but only where the robots had collected data; Helix 2.5 asks whether a humanoid can enter a home it has never seen and start work on its own.
Index pretraining is the lift: Holding task data, architecture, training and evaluation fixed, Index pretraining raised zero-shot success from 9% to 56% (full-task success only; no partial credit). Helix 2.5 used half as much task-specific data as a representative Helix 02 behaviour, then generalised across 30 homes rather than the training house.
Scaling and compute: Doubling Index pretraining data improved downstream robot-action prediction smoothly enough to forecast the largest run's loss to four decimal places. Figure says Index now generates about 35 minutes of new human experience every second, and has committed $3.5B of compute with Nscale to train Helix.