Physical Intelligence clarifies: zero-shot LLM already beats fine-tuned open VLA
paigeinsf · x · 2026-11-12
Thread post 4/4 qualifies the comparison: π0.5 needs station data to run at all, so its score reflects the 500 collected demos, and more targeted data would likely raise it; π0.5 is an older open-weight model and newer π models are likely stronger, with π0.5 having moved the field forward. The key takeaway stands: a zero-shot LLM already beats a fine-tuned open VLA on memory and long-horizon tasks — the shift worth watching.
More from Embodied
- Meta Releases XRStereo: ~18k-Sequence Synthetic Stereo Dataset on Hugging Face — _akhaliq · 2026-11-12
- Hacking the open-source Anker recorder into a real-time voice input device with Qwen 0.6B ASR — vista8 · 2026-11-12
- Liquid AI lays out roadmap to put AI on half the world's devices via model-hardware co-design — JosephJacks_ · 2026-11-12
- Wayve unveils Learned Safety Envelope: end-to-end trained live risk prediction for AI driving — alexgkendall · 2026-11-12
- Pro-0 trains dexterous robot hands from human demos with zero on-robot data — chris_j_paxton · 2026-11-12
- Chris Paxton endorses "the age of deploy cos" to bridge robotics' demo-to-deployment gap — chris_j_paxton · 2026-11-12