Real-robot eval: fine-tuned π0.5 wins box packing 51-39, zero-shot Fable dominates memory tasks
paigeinsf · x · 2026-11-12
SE3 Labs' real-robot benchmark shows a clear split: fine-tuned π0.5 scored 51 vs 39 for zero-shot Fable on box packing (stable 45-60 every run, thanks to 500 station demos), while Fable crushed memory and long-horizon tasks—64/100 vs 7 on hidden-cup ordering and 47 vs 27 on long-horizon. Zero-shot LLM planning already beats a fine-tuned open VLA where memory matters; targeted motion tasks still favor fine-tuning.
More from Embodied
- LO-World: unified BEV latent model boosts 4D occupancy +10.9% IoU, LiDAR CD -30.1% — Zhimin Liao · 2026-11-12
- CoRL 2026 to host first live Child-Robot Dexterity Challenge on stage — k7agar · 2026-11-12
- Claude Fable 5.2 tops AA-Robotics v0.1 at 75, first model to pass measure box via video-only sizing — ArtificialAnlys · 2026-11-12
- AI agents drive through GTA 6's Vice City with an army in tow — MicahBerkley · 2026-11-12
- Claude designed a humanoid robot overnight: 698 parts, 33 actuators, self-iterated balance — Distinct-Question-16 · 2026-11-12
- Meta Releases XRStereo: ~18k-Sequence Synthetic Stereo Dataset on Hugging Face — _akhaliq · 2026-11-12