Despite the existence of robots that are capable of lifting heavy loads, robotic assistants that can help people move objects as part of a team are not available. This is because of a lack of critical intelligence that results in inefficient and ineffective performance of these robots. This work makes progress towards improved intelligence of robotic lifting assistants by studying human-human teams in order to understand basic principles of co-manipulation teamwork. The effect of modus, or the manner in which a team moves an object together, is the primary study of this work. Data was collected from over 30 human-human trials in which participants in teams of two co-manipulated an object that weighed 60 pounds. These participants maneuvered through a series of five obstacles while carrying the object, exhibiting one of four modi at any given time. The raw data from these experiments was cleaned and distilled into a pose trajectory, velocity trajectory, acceleration trajectory, and interaction wrench trajectory. Classifying on the original base set of four modi with a neural net showed that two of the three modi were very similar, such that classification between three modi was more appropriate. The three modi used in classification were \emph{quickly}, \emph{smoothly} and \emph{avoiding obstacles}. Using a convolutional neural net, three modi were able to be classified from a validation set with up to 85\% accuracy. Detecting modus has the potential to greatly improve human-robot co-manipulation by providing a means to determine an appropriate robot behavior objective function. Survey data showed that participants trust each other more after working together and that they feel that their partners are more qualified after they worked together. A number of modified scales were also shown to be reliable which will allow future researchers in human-robot co-manipulation to properly evaluate how humans feel about working with each other. These same scales will also provide a useful comparison to human-robot teams in order to determine how much humans trust robots as co-manipulation team members.

When humans and remote robots work together on a team, the robots always interact with a human supervisor, even if the interaction is limited to occasional reports. Distracting a human with robotic interactions doesn't pose a problem so long as the inclusion of robots increases the team's overall effectiveness. Unfortunately, increasing the supervisor's cognitive load may decrease the team's sustainable performance to the point where robotic agents are more a liability than an asset. Present approaches resolve this problem with adaptive autonomy, where a robot changes its level of autonomy based on the supervisor's cognitive load. This thesis proposes to augment adaptive autonomy by modeling temporal latency and using this model to optimally select the temporal interval between when a supervisor is informed of a pending change and when the robot makes the change. This enables robotic team members to time their actions in response to the supervisor's cognitive load. The hypothesis is confirmed in a user-study where 26 participants interacted with a simulated search-and-rescue scenario.

Successful applications of Reinforcement Learning (RL) in the robotics field has proliferated after DeepMind and OpenAI showed the ability of RL techniques to develop intelligent robotic systems that could learn to perform complex tasks. Ever since the use of robots for surgical procedures, researchers have been trying to bring some sort of autonomy into the operating room. Surgical robotic systems such as da Vinci currently provide the surgeons with direct control. To relieve the stress and the burden on the surgeon using the da Vinci robot, semi-automating or automating surgical tasks such as suturing can be beneficial. This work presents a RL-based approach to automate the needle hand-off task. It puts forward two approaches based on the type of environment, a discrete and continuous space approach. For capturing a unique suturing style, user data was collected using the da Vinci Research Kit to generate a sparse reward function. It was used to derive an optimal policy using Q-learning for a discretized environment. Further, a RL framework for da Vinci Research Kit was developed using a real-time dynamics simulator - Asynchronous Multi-Body Framework (AMBF). A model was trained and evaluated to reach the desired goal using model-free RL techniques while considering the dynamics of the robot to help mitigate the difficulty in transferring trained model to real-world robots. Therefore, the developed RL framework would enable the RL community to train surgical robots using state of the art RL techniques and transfer it to real-world robots with minimal effort. Based on the results obtained, the viability of applying RL techniques to develop a supervised level of autonomy for performing surgical tasks is discussed. To summarize, this work mainly focuses on using RL to automate the suture hand-off task in order to move a step towards solving the greater problem of automating suturing.