VLA models encode task concepts as curved multi-dimensional geometry, 19-author study finds
ninamiolane · x · 2026-10-27
A 19-author arXiv study, 'A Phenomenology of Neural Geometry in Vision-Language-Action Models,' goes beyond linear readability to map the actual geometry of task-relevant concepts in robot VLA policies. Using controlled sweeps, natural rollouts, probes, and activation interventions, it builds a taxonomy of seven concepts: geometry is predominantly multi-dimensional and shaped by physical constraints — joint limits make wrist orientation an open arc, and 'distance to target' splits into coarse and fine rulers in two orthogonal subspaces. The same geometries recur in two more VLAs trained on real-world robot data, and steering along a concept's geometry can guide policy behavior.
More from Embodied
- Mundo turns a few room photos into an explorable, simulatable 3D world — at 1.5 FPS — mark_k · 2026-10-27
- Zurich Physical AI Hackathon puts attendees in a room with $1.5M of humanoids — ___Mufasaa · 2026-10-27
- Apple responds after iPhone Duo inner displays crack along the hinge — mobileraj · 2026-10-27
- Munro Live Posts Full Cybercab Engineering Deep-Dive Video — yunta_tsai · 2026-10-27
- Tesla's Robotaxi Fleet Keeps Growing in Austin — yunta_tsai · 2026-10-27
- Building Halo Infinite Forge maps with local LLMs, DCjson and Blender MCP — swagonflyyyy · 2026-10-27