AffordJEV: Affordance Grounding by Confirming Candidate Regions

This research presents AffordJEV, a novel approach for affordance grounding in embodied manipulation. The method employs a System One model that directly evaluates candidate regions proposed by frozen segmenters without intermediate text generation, enabling efficient inference and direct semantic control over region extents. The framework utilizes a hierarchical confirmation process from object to part levels, with bidirectional evidence passing between levels to correct decisions. Trained on the InstructPart dataset, AffordJEV outperforms existing chain-of-thought methods in domain-specific tasks and achieves better zero-shot transfer performance on UMD and IIT-AFF benchmarks than comparable System One models. The authors claim this is the first approach to apply System One models specifically to affordance grounding tasks.

Generated from attributed reports · Updated 2 days ago