Embodied Turing Machines: coding agents write reusable robot routines, 70.24% on 42 bimanual tasks
jiqizhixin · x · 2026-11-10
NTU and Ropedia's "Embodied Turing Machines" preprint has a coding agent develop reusable routines for perception, planning, and checking, then refine them via task feedback.
- A shared code library carries procedural knowledge across attempts; at execution the program reads visual and proprioceptive observations, updates state, and picks actions
- The control loop runs pure code without calling a VLM/VLA model every step; model assistance is only used to develop and revise the code
- Reports 70.24% success across 42 bimanual simulation tasks under the authors' own RoboDojo protocol
The authors note the results come from their own simulation suite, so generalization remains an open question.
More from Embodied
- Vibe coding in public: swapping Gazebo's Ogre renderer for Filament with AI — facontidavide · 2026-11-10
- UCLA's Visible Touch turns touch into image overlays, boosting VLA robot success by up to 30pp — Scobleizer · 2026-11-10
- Google DeepMind wows CoRL with Apollo humanoid taking spoken task requests from anyone — keerthanpg · 2026-11-10
- RoboCap LiDAR-100: 100 hours of egocentric LiDAR data for robots learning from humans — micoolcho · 2026-11-10
- Robot ditches feet for wheels in new demo — i_bioloid · 2026-11-10
- New-gen UMI gripper passes real-world bar test, lifting a full glass of water — YuXiang_IRVL · 2026-11-10