Figure Helix 2.5 hits 56% zero-shot success in 30 unseen homes vs 9% trained from scratch
TansuYegen · x · 2026-09-22
Figure's Helix 2.5 shows that pretraining on human behavior data dramatically improves a humanoid robot's ability to work in unfamiliar environments: an Index-pretrained model scored 56% zero-shot success across 30 unseen homes, versus just 9% for a model trained from scratch. The takeaway is that learning from broad human experience lets robots generalize instead of needing per-environment training. The author adds that if this trend holds, the hardest problem shifts from training the bot to deciding whose habits it copies.
Related event: Figure Launches Helix 2.5, Hits 56% Zero-Shot Success in 30 Unseen Homes(6 posts)→
More from Embodied
- A Perfect Chess World Model Still Won't Beat Magnus Carlsen, Researcher Argues — Machine Learning Street Talk · 2026-09-22
- Roboticist Georgia Chalathi: scale is learning to use structure, not replacing it — GeorgiaChal · 2026-09-22
- UBTECH's super factory targets 10,000 humanoids a year as revenue jumps 1,445% — lukas_m_ziegler · 2026-09-22
- A big week in robotics: Figure, Agility, Boston Dynamics and more roundup — lukas_m_ziegler · 2026-09-22
- CARE helps VLA robots recover from failures, boosting task success by up to 15.9 points — dalian-university-of-technology · 2026-09-22
- AgiBot is building an app store to swap AI personas on the same humanoid robot — Scobleizer · 2026-09-22