Oxford Robotics Institute
“A training-free execution scaffold — a predicate-gated state machine plus world-frame geometric control — turns a fine-tuned VLA that scores 0–35% on a multi-step humanoid coffee task into an 80% performer without changing a single model weight.”
Source→“the aspects a flat VLA handles least reliably, deciding when and where to act, can often be resolved directly from physical quantities, whereas learning how to execute a skill is precisely where VLAs are most effective”
Source→“PGX ran on a real Unitree G1 with Dex-3 hands performing the full coffee task — walk, grasp, place, wait, re-grasp, deliver”
Source→“GR00T N1.6 was the stronger backbone (80% coffee success inside PGX vs. 70% for π0.5), and the real robot ran "the decoupled GR00T-WBC policy" for locomotion”
Source→“50 Quest-teleoperated demos for sim coffee, 44 for the real robot (Sec. IV-A.1)”
Source→“First author on PGX and previously on KitchenVLA (iterative vision-language corrections for robot execution, cited as [38])”
Source→“Also first author of FurnitureVLA (long-horizon bimanual furniture assembly, cited as [13])”
Source→“Director, Oxford Robotics Institute, University of Oxford”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.