Visual Foresight With a Local Dynamics Model

TL;DR

This paper introduces the Local Dynamics Model (LDM), a sample-efficient approach for learning state transitions in manipulation tasks, enabling effective one-step lookahead planning that outperforms existing methods.

Contribution

The paper presents the LDM, a novel model architecture that improves sample efficiency and planning performance in manipulation tasks compared to prior models.

Findings

LDM is more sample-efficient than other architectures.

Combining LDM with planning outperforms other policies.

LDM achieves superior results in simulation manipulation tasks.

Abstract

Model-free policy learning has been shown to be capable of learning manipulation policies which can solve long-time horizon tasks using single-step manipulation primitives. However, training these policies is a time-consuming process requiring large amounts of data. We propose the Local Dynamics Model (LDM) which efficiently learns the state-transition function for these manipulation primitives. By combining the LDM with model-free policy learning, we can learn policies which can solve complex manipulation tasks using one-step lookahead planning. We show that the LDM is both more sample-efficient and outperforms other model architectures. When combined with planning, we can outperform other model-based and model-free policies on several challenging manipulation tasks in simulation.