A comparison of contact distribution representations for learning to predict object interactions

Different contacts between objects afford different interactions between them. For example, while contacts below an object can provide support, contacts on opposing sides can be used for pinching. Hence, a robot can learn to predict which interactions are currently afforded based on the set of contacts. However, representing sets of contacts is not trivial, as the number of contacts is not fixed nor are the contacts ordered. In this paper, we compare different methods for representing contacts, including bag-of-features, probability product kernels, and random forests. These approaches model the distribution over the contacts without relying on task-specific features. The methods were evaluated on both simulated grasping data, as well as real robot grasps. The random forest and the normalized expected likelihood kernel approaches achieved the highest accuracies for the simulated experiments. In the case of the real robot data, the more robust exponential χ2 and Bhattacharyya kernels achieved higher accuracies.