Relationship Oriented Affordance Learning through Manipulation Graph Construction
This addresses the problem of enabling robots to understand and manipulate objects in arbitrary scenes, which is incremental as it builds on existing affordance learning methods.
The paper tackles the problem of learning manipulation affordances by proposing a Manipulation Relationship Graph (MRG) representation and AR-Net model, achieving success rates of 88.89% on task relationship recognition and 73.33% on task completion.
In this paper, we propose Manipulation Relationship Graph (MRG), a novel affordance representation which captures the underlying manipulation relationships of an arbitrary scene. To construct such a graph from raw visual observations, a deep nerual network named AR-Net is introduced. It consists of an Attribute module and a Context module, which guide the relationship learning at object and subgraph level respectively. We quantitatively validate our method on a novel manipulation relationship dataset named SMRD. To evaluate the performance of the proposed model and representation, both visual perception and physical manipulation experiments are conducted. Overall, AR-Net along with MRG outperforms all baselines, achieving the success rate of 88.89% on task relationship recognition (TRR) and 73.33% on task completion (TC)