SOLAR: Deep Structured Representations for Model-Based Reinforcement Learning
This addresses the problem of data efficiency and complexity in model-based RL for robotics with image observations, representing an incremental improvement by adapting existing methods to new domains.
The paper tackles the challenge of applying model-based reinforcement learning to domains with complex image observations by learning deep structured representations optimized for inferring simple dynamics and cost models, enabling the use of linear-quadratic regulator methods on robotics tasks and achieving substantially better final performance than other model-based methods while being more efficient than model-free RL.
Model-based reinforcement learning (RL) has proven to be a data efficient approach for learning control tasks but is difficult to utilize in domains with complex observations such as images. In this paper, we present a method for learning representations that are suitable for iterative model-based policy improvement, even when the underlying dynamical system has complex dynamics and image observations, in that these representations are optimized for inferring simple dynamics and cost models given data from the current policy. This enables a model-based RL method based on the linear-quadratic regulator (LQR) to be used for systems with image observations. We evaluate our approach on a range of robotics tasks, including manipulation with a real-world robotic arm directly from images. We find that our method produces substantially better final performance than other model-based RL methods while being significantly more efficient than model-free RL.