A human-centered safe robot reinforcement learning framework with interactive behaviors.

Deployment of Reinforcement Learning (RL) algorithms for robotics applications in the real world requires ensuring the safety of the robot and its environment. Safe Robot RL (SRRL) is a crucial step toward achieving human-robot coexistence. In this paper, we envision a human-centered SRRL framework consisting of three stages: safe exploration, safety value alignment, and safe collaboration. We examine the research gaps in these areas and propose to leverage interactive behaviors for SRRL. Interactive behaviors enable bi-directional information transfer between humans and robots, such as conversational robot ChatGPT. We argue that interactive behaviors need further attention from the SRRL community. We discuss four open challenges related to the robustness, efficiency, transparency, and adaptability of SRRL with interactive behaviors.

Dataset of bimanual human-to-human object handovers.

We present a multi-sensor dataset of bimanual human-to-human object handovers. The dataset consists of 240 recordings obtained from 12 pairs of participants performing bimanual object handovers with 10 objects, and 120 recordings obtained from the same 12 pairs of participants performing unimanual handovers with 5 of those objects. Each recording includes the giver and receiver's 13 upper-body bone position and orientation trajectories, position trajectories for the 27 markers placed on their upper bodies, object position and orientation trajectories, and two RGB-D data streams. The motion trajectories are recorded at 120Hz and the RGB-D streams are recorded at 30Hz. The recordings are annotated with the three handover phases: reach, transfer, and retreat. The dataset also includes four anthropometric measurements of the participants: height, waistline height, arm span, and weight. Our dataset could help investigations of the bimanual reaching motions and grasps utilized by humans while performing handovers. Also, it can be used to train robots to perform bimanual object handovers with humans.

Human Preferences for Robot Eye Gaze in Human-to-Robot Handovers.

This paper investigates human's preferences for a robot's eye gaze behavior during human-to-robot handovers. We studied gaze patterns for all three phases of the handover process: reach, transfer, and retreat, as opposed to previous work which only focused on the reaching phase. Additionally, we investigated whether the object's size or fragility or the human's posture affect the human's preferences for the robot gaze. A public data-set of human-human handovers was analyzed to obtain the most frequent gaze behaviors that human receivers perform. These were then used to program the robot's receiver gaze behaviors. In two sets of user studies (video and in-person), a collaborative robot exhibited these gaze behaviors while receiving an object from a human. In the video studies, 72 participants watched and compared videos of handovers between a human actor and a robot demonstrating each of the three gaze behaviors. In the in-person studies, a different set of 72 participants physically performed object handovers with the robot and evaluated their perception of the handovers for the robot's different gaze behaviors. Results showed that, for both observers and participants in a handover, when the robot exhibited Face-Hand-Face gaze (gazing at the giver's face and then at the giver's hand during the reach phase and back at the giver's face during the retreat phase), participants considered the handover to be more likable, anthropomorphic, and communicative of timing ( p < 0.0001 ) . 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.