Generalizing Object-Centric Task-Axes Controllers using Keypoints
This addresses the problem of robotic manipulation generalization for real-world applications, but it appears incremental as it builds on modular and object-centric approaches.
The paper tackles the challenge of robots manipulating objects with diverse shapes and sizes without geometric models by proposing modular task policies that compose object-centric task-axes controllers, showing generalization to large variance in object properties across multiple manipulation tasks.
To perform manipulation tasks in the real world, robots need to operate on objects with various shapes, sizes and without access to geometric models. It is often unfeasible to train monolithic neural network policies across such large variance in object properties. Towards this generalization challenge, we propose to learn modular task policies which compose object-centric task-axes controllers. These task-axes controllers are parameterized by properties associated with underlying objects in the scene. We infer these controller parameters directly from visual input using multi-view dense correspondence learning. Our overall approach provides a simple, modular and yet powerful framework for learning manipulation tasks. We empirically evaluate our approach on multiple different manipulation tasks and show its ability to generalize to large variance in object size, shape and geometry.