Imagination Policy: Using Generative Point Cloud Models for Learning Manipulation Policies

TL;DR

This paper introduces Imagination Policy, a novel approach that uses generative point cloud models to imagine goal states for high-precision manipulation, achieving high efficiency and generalization in robotic pick-and-place tasks.

Contribution

It presents a new multi-task key-frame policy network that generates point clouds to imagine goal states, transforming action inference into a local generative task for robotic manipulation.

Findings

Achieves state-of-the-art performance on RLbench benchmark

Demonstrates high sample efficiency and generalization

Validates effectiveness on real robot experiments

Abstract

Humans can imagine goal states during planning and perform actions to match those goals. In this work, we propose Imagination Policy, a novel multi-task key-frame policy network for solving high-precision pick and place tasks. Instead of learning actions directly, Imagination Policy generates point clouds to imagine desired states which are then translated to actions using rigid action estimation. This transforms action inference into a local generative task. We leverage pick and place symmetries underlying the tasks in the generation process and achieve extremely high sample efficiency and generalizability to unseen configurations. Finally, we demonstrate state-of-the-art performance across various tasks on the RLbench benchmark compared with several strong baselines and validate our approach on a real robot.