Manipulate-Anything: VLMs auto-generate real-world robot demo data, beating VoxPoser
DJiafei · x · 2026-11-18
Researchers from UW, NVIDIA and AI2 present MANIPULATE-ANYTHING, a scalable method that uses vision-language models to automatically generate demonstration data for real-world robotic manipulation.
- Needs no privileged state info or hand-designed skills; can manipulate any static object and solve tasks zero-shot
- Generated successful trajectories for all 7 real-world and 14 simulation tasks, significantly outperforming VoxPoser, Code as Policies and Scaling Up
- Behavior cloning policies trained on its data are more robust than those trained on human demos or competitor-generated data
- Paper, code and generated data are open-sourced
More from Embodied
- Tokkey debuts thumb-sized AI hardware to run local models on your Mac — TejasKumar_ · 2026-11-18
- 'Nothing more physical AI than moving tons of rock': live demo at Web Summit shows real-world difficulty — MarvinTBaumann · 2026-11-18
- StellarEdge Auriga report: verifiable execution harness hits 57.74% on 42 RoboDojo tasks — jiqizhixin · 2026-11-18
- Japanese team tests sidewalk delivery robot in public-road trials in Okayama — 4310sy · 2026-11-18
- Shanghai AI Lab open-sources EgoExplorer, a wearable pipeline for whole-body mobile manipulation data — 量子位 · 2026-11-18
- Tesla Cybercab sightings hit 300 on the road; 200 new Texas registrations, 773 total — JOBhakdi · 2026-11-18