For a Robot, a Drawer Must Open, Not Just Be Seen: Building 10,000 Interactive Object Assets
Blog post from Superb AI
As part of Korea’s Dokpamo Sovereign AI Foundation Model Project, Phase 2 converted imagery from 300,000 real Korean home-environment frames into 10,000 interactive household-object assets intended to train robots for manipulation tasks. The assets include pixel-level segmentation and, for eight articulated categories such as cabinets, refrigerators, drawers, wardrobes, and curtains, mechanical properties including motion axes and ranges; rigid objects such as cups and plates were modeled separately. An AI-assisted workflow used SAM and computer-vision methods for initial labeling, a vision-language model for review, and human reviewers for final quality control, applying a dataset-wide passing threshold of at least 0.7 IoU and 90% semantic accuracy. Occlusion in realistic household scenes was identified as the principal segmentation challenge, while curtains required a simplified sliding-motion model to balance realism with computational cost. The project positions its focus on Korean home contexts and articulated objects alongside related efforts such as Meta’s SAM 3D and AgiBot World, with a future phase planned to combine spatial, object, and behavior assets into a synthetic-data pipeline.
No tracked trend matches for this post yet.
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.