Figure's Helix 2.5: zero-shot whole-body autonomy in 30 unseen homes (Sept 17, 2026)
AI-assisted source briefing. Figure reports Helix 2.5, a foundation-model–based humanoid policy pretrained on its Index dataset, that performed three long-horizon whole-body tasks zero-shot across 30 Bay Area homes. The company attributes most gains to Index pretraining and reports a human-to-humanoid data-s
Published by The Humanoid Race
Source link: www.figure.ai

AI-assisted source briefing. Automatically written and checked against the cited source, without human editorial review. These are automated checks, not independent verification. Evidence date: 2026-09-17. Classification: reported actual. Source date: September 17, 2026. Figure says Helix 2.5 was evaluated zero‑shot across 30 Bay Area homes: the robot performed three long‑horizon whole‑body behaviors (living‑room tidying, towel folding, and bed making) in homes and with objects the system had never seen, with no environment data collected in any evaluation home. (Figure’s claim and evaluation framing are described in the announcement.) How the system was built and tested: Figure says Helix 2.5 was pretrained on Index, its large, human‑behavior dataset, then adapted from that single foundation model into three distinct whole‑body behaviors spanning locomotion, rigid and deformable manipulation, bimanual coordination, and active perception. Each behavior used a single fixed checkpoint across all 30 homes and no weights were adapted to evaluation homes or objects. Quantified impact of Index pretraining: In a controlled comparison where task data, architecture, and training were held fixed, Figure reports the model initialized from Index succeeded on 56% of zero‑shot trials while a policy trained from random initialization succeeded on 9% — a greater than sixfold difference. Figure states that success required completing entire tasks end‑to‑end (no partial credit). Data efficiency and recovery behaviors: Figure reports Helix 2.5 needed roughly half as much task‑specific adaptation data as a representative Helix 02 behavior to match its success rate, while generalizing zero‑shot across 30 unseen homes. The company also highlights improved whole‑body self‑correction (e.g., stepping back, changing stance, moving around obstacles) as a qualitative effect they attribute to Index pretraining. Scaling and compute claims: Figure reports a measured human‑to‑humanoid transfer scaling law: training on nested subsets of Index (an 8× pretraining data range) produced predictable reductions in downstream action‑prediction loss, allowing them to forecast their largest run’s loss to four decimal places; they say forecasting error was 0.54% of the variation across the range. Figure also states Index is generating roughly 35 minutes of new human experience every second and that they committed $3.5B of compute to training Helix. Evaluation protocol and success criteria: Figure describes precommitment of trial rubrics and blind evaluation procedures: initial configurations were unique per trial, objects used in evaluation were verified not to appear in task‑specification data, and trials were graded against fixed pass/fail criteria (e.g., all 13–15 toys placed in a basket for Living‑Room Tidy; towels folded and placed in basket; pillows and comforter corners placed at top third of bed with comforter smoothed). Timeouts, per‑item grading rules, and abort conditions were defined ahead of evaluation. Original source: Helix 2.5: Zero-Shot 30-Home Generalization