Figure's Helix 2.5 hit 56% success in 30 unseen homes, vs 9% from scratch
CyberRobooo · x · 2026-09-18
Figure's Helix 2.5 shows how human data helps robots generalize:
- F.03 humanoids did household tasks in 30 rented Bay Area homes they had never seen, with no new data collected there.
- With large-scale human behavior pretraining, success rate was 56% vs just 9% for a model trained from zero shot — implying the earlier Helix 02 demo's real success rate was only 10%.
- Figure found 8× more human pretraining data kept reducing action-prediction loss, suggesting robot capability may scale with human experience much like foundation models.
The takeaway: robot capability may hinge on pretraining data scale (tens to hundreds of millions of hours), and the golden age of human behavior data collection is here.
Related event: Figure Unveils Helix 2.5: Zero-Shot Household Work Across 30 Unseen Homes(27 posts)→
More from Embodied
- Autonomous's Harness Pro uses an ESP32-P4, not a mini PC, to monitor AI coding agents — dee_hw · 2026-11-02
- KAIST's QCP lets robot policies control execution quality in real time — kaist-ai · 2026-11-02
- ANYbotics begins customer deliveries of ANYmal X, first legged robot certified for explosive atmospheres — lukas_m_ziegler · 2026-11-02
- InSpatio opens China's first embodied simulation data center targeting 1M hours/month — 智东西 · 2026-11-02
- GR-BC mines better trajectories from imperfect demos, hitting ~91% success on precise manipulation — 机器之心 · 2026-11-02
- RAM prices tripled, so Tesla cut a third of memory from Optimus's brain — aakashgupta · 2026-11-02