Stanford researchers build kitchen robot with 89% accuracy
New robotics pipeline handles cluttered kitchens without retraining...
A new paper presents a modular perception pipeline for kitchen robotic manipulation, focusing on dishware handling. It combines open-vocabulary detection, multi-view segmentation, instance-aware 3D reconstruction, and 2D-3D feature fusion for 6D pose estimation and grasp planning. The best configuration—LLMDet + SAMv2 + DINOv2 + GeoTransformer—achieves an ADI of 89.12% on a 20-scene kitchen benchmark with cluttered and occluded conditions. Real-world tests confirm it can run on physical robots without environment-specific retraining, successfully performing sink-to-dishwasher transfer and cup stacking. Code and supplementary materials are linked in the paper.
- Combines LLMDet + SAMv2 + DINOv2 + GeoTransformer foundation models for kitchen robotics
- Achieved 89.12% accuracy on cluttered kitchen benchmarks with zero-shot deployment
- Successfully executed real-world tasks like dishwasher transfers and cup stacking
Why It Matters
Proves autonomous household robots can now operate in unstructured environments without custom retraining, accelerating real-world deployment.