Figure shows pretraining scaling works for robots: 9% to 56% success in unseen homes

量子位 · wechat · 2026-09-23

Figure put Helix 2.5 into 30 unseen Bay Area homes to tidy toys, fold towels and make beds. 'Zero-shot' means no per-home retraining — the tasks were previously adapted on data from other environments.

Two key experiments: with identical architecture and downstream data, random-init policies scored 9% full-task success vs 56% after Index (human-behavior data) pretraining; and nested pretraining data spanning 8x scale showed predictable loss reduction, with a fitting error of just 0.54%. Note this second result is an offline metric — Figure hasn't published the curve from pretraining scale to real-home success.

Wider landscape: Dyna-2 scaled first-person human video from 1k to 1M hours, lifting mean normalized scores on 14 real tasks from 20% to 53% — scaling reaching closed-loop, with ablations showing a video-prediction objective is needed for monotonic gains. Light-O1 (LightOrigins) pretrained on internet-scale third-person human video (3.75B→120B tokens), reporting power-law error decay it calls a 'Transfer Scaling Law', still open-loop.

Figure's playbook: a Creator network has collected 16M+ videos and paid out $15M, with new data arriving at 35 minutes per second; plus a $3.5B (expandable to $6B+) compute deal with Nscale for up to 100k NVIDIA GPUs. Alternative routes (CoMo, HuRo, Rhoda) try to extract transferable signal from internet video instead.

Original post →

Figure shows pretraining scaling works for robots: 9% to 56% success in unseen homes

More from Embodied