VLBiMan++ learns pouring, tool use and folding from a single human demo, no retraining
stepjamUK · x · 2026-09-17
VLBiMan++, a bimanual robot system from Shenzhen University and DexForce, cuts task learning from hundreds of demonstrations down to one.
- Conventional robots need hundreds of demos per task and hundreds more whenever the task changes. VLBiMan++ watches a single human pouring demo, decomposes it into reusable sub-skills, then uses vision-language grounding to adapt those skills as situations change — no retraining, extra demos, or fine-tuning.
- The range of "changes" is the highlight: the same one-shot prior transfers across pouring, reorienting and unscrewing caps, spoon and funnel tool use, zipping a pen bag, coiling cables, and folding towels — rigid, articulated, and deformable objects all handled by one approach.
The key signal is cross-task generalization, not single-task success rates.
More from Embodied
- Princeton's Intro to Robotics Course Videos, Slides and Code Now Free Online — Majumdar_Ani · 2026-09-17
- Critic slams humanoid robot engineering culture: a hundred firms building rote-trained "mechanical drones" — generativist · 2026-09-17
- Open-source Openclaw gives humanoids spatial-temporal understanding, works on Unitree G1 — hackgoofer · 2026-09-17
- BCI and computer architecture expert Bhattacharjee joins Princeton CS from Yale — LakeBrenden · 2026-09-17
- Japan's Ugo unveils Nova, a wheeled dual-arm robot with 10kg payload for factories — CyberRobooo · 2026-09-17
- Baidu's Apollo Go begins fully driverless testing on Hong Kong roads — Baidu_Inc · 2026-09-17