HUMAN ROBOT HANDOVERS

investigates the collaborative task of human to robot object handovers. Handovers are a vital capability for collaborative robots. We focused on two crucial issues for embedding humanlike characteristics into robots. First, we examined the impact of robot's non-verbal communication on human's experience and fluency of human to robot handovers. Second, we developed and evaluated a robot controller based on reinforcement learning to perform a more natural sequential handover.

The first part in the research investigated human's preference of the robot's eye gaze during humanrobot handovers. While there is some literature on robot gaze in robot-to-human handovers, there is a dearth of literature on robot gaze in human-to-robot handovers. Prior research that studied robot gaze behavior in human-to-robot handovers considered only the receiver's gaze patterns in the "reach" phase and used only one particular object in one configuration. Building upon this work, this research studied gaze patterns for all three phases of the handover process: reach, transfer, and retreat, both in video and in-person studies. This included investigation of whether the object's size and fragility or the giver's posture affect the human's preference of the robot gaze in terms of the perceived liking, anthropomorphism, and timing communication of the handover.

A public data-set of handovers videos were analyzed frame-by-frame to determine the most frequent gaze behaviors in human-human handovers. The most frequent gaze behaviors were found to be: gazing at the giver's hand and then at the giver's face (Hand-Face gaze), gazing initially at the giver's face and then at the giver's hand and then back to look at the giver's face (Face-Hand-Face gaze), and continuously look at the giver's hand (Hand gaze).

A Sawyer collaborative 7 DOF (degrees of freedom) robot was programmed to perform the handover task and exhibit these gaze behaviors. Different objects with different types of giverreceiver configurations were analyzed in two studies – a video study and an in-person study. In the video study, 72 participants watched and compared videos of human to robot handovers between an actor and a robot demonstrating the three gaze behaviors. In the in-person study, a different set of participants physically performed object handovers with the robot and evaluated their perception of the handovers for the robot's different gaze behaviors. Results revealed that for both studies when the robot initially gazes at the giver's face and then at the giver's hand and then back at the giver's face (Face-Hand-Face gaze), participants consider the handover to be more likable, anthropomorphic, and communicative of timing (p < 0.005). However, we did not find evidence of any effect of the object's size or fragility or the giver's posture on the gaze preference.

In the second part of the research, we assessed the potential of a model-based reinforcement learning (RL) method, the Guided Policy Search (GPS), to train a robot controller for human-robot object handovers. GPS is a data-efficient system that does not necessitate prior knowledge of the robot and environment dynamics, providing a promising approach for the handover task. Nevertheless, despite GPS demonstration on various navigation tasks and autonomous manipulation, testing GPS in a physical human-robot collaborative task has not been reported. In this study, the reach phase of a handover is formulated as an RL problem, with subsequent training of the Panda collaborative 7 DOF robot arm both in a simulation environment and directly on the physical robot.

Our results indicate that testing the policy learnt in the simulation environment on the real robot, is an infeasible solution for real world implementation. When estimating only static targets, we found that the performance of the global policies learnt by GPS generalize relatively well. However, the global policy performance got slightly improved by adding local controllers in regions with highest test errors. When evaluating the global policy trained with static targets on a moving target, the robot generated highly inefficient trajectories and reached areas outside of its cartesian position limits. Training on moving targets improved trajectories, but resulted with significantly larger worst-case errors. However, this issue can be addressed by adding local controllers to the training phase, improving the global policy’s performance