Fable 5.2 robot evals: overall success rate jumps from 34% to 50% across 456 tasks
GuanyaShi · x · 2026-11-13
The Embodied First Exam team benchmarked the newly released Fable 5.2 on 456 core robot tasks against Astra (all Standard Mode):
- Overall success rate: 34% → 50%
- BEHAVIOR-1K (long-horizon mobile manipulation): 20% → 37%
- iSpatialGym (spatial reasoning + manipulation): 50% → 88%
- Embodied Olympics-Fun (dynamic tasks, whole-body control): 36% → 48%
- DexHandExam (dexterous manipulation): 57% → 65%
- Embodied Safety Exam: 64% → 70%
- RoboPaint (painting): 32% → 9% — Astra remains the better artist
A detailed analysis is coming soon; the full leaderboard is already live.
More from Embodied
- Ukrainian-founded drone startup Uforce raises $700M at $6B valuation — Polymarket · 2026-11-13
- With 5 tradespeople retiring for every 2 entering, Mentra bets AI glasses can close the skills gap — caydengineer · 2026-11-13
- CoRL workshops overflow: attendees on floors and hallways at continual learning session — DominiqueCAPaul · 2026-11-13
- UW Prof. Abhishek Gupta muses on academia's role in robotics RL research at CSIR workshop — chris_j_paxton · 2026-11-13
- YC F26 startup Kairox builds autonomous drones that both fly and swim — ycombinator · 2026-11-13
- Mila brings its robotics research to CoRL 2026 in Austin — Mila_Quebec · 2026-11-13