Matching Items (3)
Description
In the field of machine learning, reinforcement learning stands out for its ability to explore approaches to complex, high dimensional problems that outperform even expert humans. For robotic locomotion tasks reinforcement learning provides an approach to solving them without the need for unique controllers. In this thesis, two reinforcement learning algorithms, Deep Deterministic Policy Gradient and Group Factor Policy Search are compared based upon their performance in the bipedal walking environment provided by OpenAI gym. These algorithms are evaluated on their performance in the environment and their sample efficiency.
ContributorsMcDonald, Dax (Author) / Ben Amor, Heni (Thesis director) / Yang, Yezhou (Committee member) / Barrett, The Honors College (Contributor) / Computer Science and Engineering Program (Contributor)
Created2018-12
Description
As technological advancements in silicon, sensors, and actuation continue, the development of robotic swarms is shifting from the domain of science fiction to reality. Many swarm applications, such as environmental monitoring, precision agriculture, disaster response, and lunar prospecting, will require controlling numerous robots with limited capabilities and information to redistribute among multiple states, such as spatial locations or tasks. A scalable control approach is to program the robots with stochastic control policies such that the robot population in each state evolves according to a mean-field model, which is independent of the number and identities of the robots. Using this model, the control policies can be designed to stabilize the swarm to the target distribution. To avoid the need to reprogram the robots for different target distributions, the robot control policies can be defined to depend only on the presence of a “leader” agent, whose control policy is designed to guide the swarm to a particular distribution. This dissertation presents a novel deep reinforcement learning (deep RL) approach to designing control policies that redistribute a swarm as quickly as possible over a strongly connected graph, according to a mean-field model in the form of the discrete-time Kolmogorov forward equation. In the leader-based strategies, the leader determines its next action based on its observations of robot populations and shepherds the swarm over the graph by probabilistically repelling nearby robots. The scalability of this approach with the swarm size is demonstrated with leader control policies that are designed using two tabular Temporal-Difference learning algorithms, trained on a discretization of the swarm distribution. To improve the scalability of the approach with robot population and graph size, control policies for both leader-based and leaderless strategies are designed using an actor-critic deep RL method that is trained on the swarm distribution predicted by the mean-field model. In the leaderless strategy, the robots’ control policies depend only on their local measurements of nearby robot populations. The control approaches are validated for different graph and swarm sizes in numerical simulations, 3D robot simulations, and experiments on a multi-robot testbed.
ContributorsKakish, Zahi Mousa (Author) / Berman, Spring (Thesis advisor) / Yong, Sze Zheng (Committee member) / Marvi, Hamid (Committee member) / Pavlic, Theodore (Committee member) / Pratt, Stephen (Committee member) / Ben Amor, Hani (Committee member) / Arizona State University (Publisher)
Created2021
Description
Autonomous Driving (AD) systems are being researched and developed actively in recent days to solve the task of controlling the vehicles safely without human intervention. One method to solve such task is through deep Reinforcement Learning (RL) approach. In deep RL, the main objective is to find an optimal control behavior, often called policy performed by an agent, which is AD system in this case. This policy is usually learned through Deep Neural Networks (DNNs) based on the observations that the agent perceives along with rewards feedback received from environment.However, recent studies demonstrated the vulnerability of such control policies learned through deep RL against adversarial attacks. This raises concerns about the application of such policies to risk-sensitive tasks like AD. Previous adversarial attacks assume that the threats can be broadly realized in two ways: First one is targeted attacks through manipu- lation of the agent’s complete observation in real time and the other is untargeted attacks through manipulation of objects in environment. The former assumes full access to the agent’s observations at almost all time, while the latter has no control over outcomes of attack. This research investigates the feasibility of targeted attacks through physical adver- sarial objects in the environment, a threat that combines the effectiveness and practicality.
Through simulations on one of the popular AD systems, it is demonstrated that a fixed optimal policy can be malfunctioned over time by an attacker e.g., performing an unintended self-parking, when an adversarial object is present. The proposed approach is formulated in such a way that the attacker can learn a dynamics of the environment and also utilizes common knowledge of agent’s dynamics to realize the attack. Further, several experiments are conducted to show the effectiveness of the proposed attack on different driving scenarios empirically. Lastly, this work also studies robustness of object location, and trade-off between the attack strength and attack length based on proposed evaluation metrics.
ContributorsBuddareddygari, Prasanth (Author) / Yang, Yezhou (Thesis advisor) / Ren, Yi (Committee member) / Fainekos, Georgios (Committee member) / Arizona State University (Publisher)
Created2021