We introduce a framework that combines simple and complex continuous state-action simulators with a real-world robot to efficiently find good control policies, while minimizing the number of samples needed from the physical robot. The framework combines the strengths of various simulation levels by first finding optimal policies in a simple model, and then using that solution to initialize a gradient-based learner in a more complex simulation. The policy and transition dynamics from the complex simulation are in turn used to guide the learning in the physical world. A method is developed for transferring information gathered in the physical world back to the learning agent in the simulation. The new information is used to re-evaluate whether the original simulated policy is still optimal given the updated knowledge from the real-world. This reverse transfer is critical to minimizing samples from the physical world. The new framework is demonstrated on a robotic car learning to perform controlled drifting maneuvers. A video of the car's performance can be found at https: //youtu.be/opsmd5yuBF0.

This paper presents a multiagent path planning algorithm based on sequential convex programming (SCP) that finds locally optimal trajectories. Previous work using SCP efficiently computes motion plans in convex spaces with no static obstacles. In many scenarios where the spaces are non-convex, previous SCP-based algorithms failed to find feasible solutions because the convex approximation of collision constraints leads to forming a sequence of infeasible optimization problems. This paper addresses this problem by tightening collision constraints incrementally, thus forming a sequence of more relaxed, feasible intermediate optimization problems. We show that the proposed algorithm increases the probability of finding feasible trajectories by 33% for teams of more than three vehicles in non-convex environments. Further, we show that decoupling the multiagent optimization problem to a number of single-agent optimization problems leads to significant improvement in computational tractability. We develop a decoupled implementation of the proposed algorithm, abbreviated dec-iSCP. We show that dec-iSCP runs 14% faster and finds feasible trajectories with higher probability than a decoupled implementation of previous SCP-based algorithms. The proposed algorithm is real-time implementable and is validated through hardware experiments on a team of quadrotors.

Autonomous learning through interaction with the physical world is a promising approach to designing controllers and decision-making policies for robots. Unfortunately, learning on robots is often difficult due to the large number of samples needed for many learning algorithms. Simulators are one way to decrease the samples needed from the robot by incorporating prior knowledge of the dynamics into the learning algorithm. In this paper we present a novel method for transferring data from a simulator to a robot, using simulated data as a prior for real-world learning. A Bayesian nonparametric prior is learned from a potentially black-box simulator. The mean of this function is used as a prior for the Probabilistic Inference for Learning Control (PILCO) algorithm. The simulated prior improves the convergence rate and performance of PILCO by directing the policy search in areas of the state-space that have not yet been observed by the robot. Simulated and hardware results show the benefits of using the prior knowledge in the learning framework.

Reinforcement learning (RL) can be a tool for designing policies and controllers for robotic systems. However, the cost of real-world samples remains prohibitive as many RL algorithms require a large number of samples before learning useful policies. Simulators are one way to decrease the number of required real-world samples, but imperfect models make deciding when and how to trust samples from a simulator difficult. We present a framework for efficient RL in a scenario where multiple simulators of a target task are available, each with varying levels of fidelity. The framework is designed to limit the number of samples used in each successively higher-fidelity/cost simulator by allowing a learning agent to choose to run trajectories at the lowest level simulator that will still provide it with useful information. Theoretical proofs of the framework's sample complexity are given and empirical results are demonstrated on a remote-controlled car with multiple simulators. The approach enables RL algorithms to find near-optimal policies in a physical robot domain with fewer expensive real-world samples than previous transfer approaches or learning without simulators.

We present a framework for reinforcement learning (RL) in a scenario where multiple simulators are available with decreasing amounts of fidelity to the real-world learning scenario. Our framework is designed to limit the number of samples used in each successively higher-fidelity/cost simulator by allowing the agent to choose to run trajectories at the lowest level that will still provide it with information. The approach transfers state-action Q-values from lower-fidelity models as heuristics for the “Knows What It Knows” family of RL algorithms, which is applicable over a wide range of possible dynamics and reward representations. Theoretical proofs of the framework's sample complexity are given and empirical results are demonstrated on a remote controlled car with multiple simulators. The approach allows RL algorithms to find near-optimal policies for the real world with fewer expensive real-world samples than previous transfer approaches or learning without simulators.

Commonly used Proportional-Integral-Derivative based UAV flight controllers are often seen to provide adequate trajectory-tracking performance, but only after extensive tuning. The gains of these controllers are tuned to particular platforms, which makes transferring controllers from one UAV to other time-intensive. This paper formulates the problem of control-transfer from a source system to a transfer system and proposes a solution that leverages well-studied techniques in adaptive control. It is shown that concurrent learning adaptive controllers improve the trajectory tracking performance of a quadrotor with the baseline linear controller directly imported from another quadrotor whose inertial characteristics and throttle mapping are very different. Extensive flight-testing, using indoor quadrotor platforms operated in MIT's RAVEN environment, is used to validate the method.

Reward learning from demonstration is the task of inferring the intents or goals of an agent demonstrating a task. Inverse reinforcement learning methods utilize the Markov decision process (MDP) framework to learn rewards, but typically scale poorly since they rely on the calculation of optimal value functions. Several key modifications are made to a previously developed Bayesian nonparametric inverse reinforcement learning algorithm that avoid calculation of an optimal value function and no longer require discretization of the state or action spaces. Experimental results given demonstrate the ability of the resulting algorithm to scale to larger problems and learn in domains with continuous demonstrations.

This video submission presents a design concept of an autonomous variable-pitch quadrotor with constant motor speed. The main aim of this work is to increase the maneuverability of the quadrotor vehicle concept while largely maintaining its mechanical simplicity. This added maneuverability will allow autonomous agile maneuvers like inverted hover and flip. A custom in lab built quadrotor with onboard attitude stabilization is developed and tested in the ACL's (Aerospace Controls Laboratory) RAVEN (Real-time indoor Autonomous Vehicle test ENvironment). Initial flight results show that the quadrotor is capable of waypoint tracking and hovering both upright and inverted.