Data-Efficient Learning for Complex and Real-Time Physical Problem Solving Using Augmented Simulation

Humans quickly solve tasks in novel systems with complex dynamics, without requiring much interaction. While deep reinforcement learning algorithms have achieved tremendous success in many complex tasks, these algorithms need a large number of samples to learn meaningful policies. In this letter, we present a task for navigating a marble to the center of a circular maze. While this system is very intuitive and easy for humans to solve, it can be very difficult and inefficient for standard reinforcement learning algorithms to learn meaningful policies. We present a model that learns to move a marble in the complex environment within minutes of interacting with the real system. Learning consists of initializing a physics engine with parameters estimated using data from the real system. The error in the physics engine is then corrected using Gaussian process regression, which is used to model the residual between real observations and physics engine simulations. The physics engine augmented with the residual model is then used to control the marble in the maze environment using a model-predictive feedback over a receding horizon. To the best of our knowledge, this is the first time that a hybrid model consisting of a full physics engine along with a statistical function approximator has been used to control a complex physical system in real-time using nonlinear model-predictive control (NMPC).

I. Introduction

Artificial Intelligence has long had the goal of designing robotic agents that can interact with the (complex) physical world in flexible, data-efficient and generalizable ways [1], [2]. Model-based control methods form plans based on predefined models of the world dynamics. However, although data-efficient, these systems require accurate dynamics models, which may not exist for complex tasks. Model-free methods on the other hand rely on reinforcement learning, where the agents simultaneously learn a model of the world dynamics and a control policy [3], [4]. However, although these methods can learn policies to solve tasks involving complex dynamics, training these policies is inefficient, as they require many samples. Furthermore, these method are typically not generalizable beyond the trained scenarios.