Usability Study of VR Interfaces for Learning from Demonstrations in Bimanual Tasks
Keywords: Learning from Demonstrations, Teleoperation Interfaces, Bimanual Tasks, Behavior cloning, Virtual Reality
TL;DR: Comparison of Two VR Interfaces for collecting demonstrations for Learning from Demonstrations in Bimanual Tasks.
Abstract: Learning from demonstrations (LfD) is a key component of state-of-the-art robot learning approaches that enables
robots to learn complex tasks by observing and imitating human
actions. While there is a large body work focused on developing
effective algorithms, demonstration quality remains a bottleneck
in LfD, mostly due to suboptimal interfaces for collecting
demonstrations. This paper addresses this gap specifically in
the context of bimanual tasks by proposing a VR setup for
demonstration data collection in which we compare two conditions: one in which the user teleoperates robot with the robot
always visible (teleoperation condition), and another where the
user demonstrates the task independently without visual feedback
(egocentric condition). The task involves two Panda robot arms
working collaboratively to pick up a tray stacked with cubes
and place it at a designated goal. Performance is measured
based on success rate and completion time. Additionally, we
conducted a user study to evaluate the user experience within VR
environments. The collected data was then fed into a behavior
cloning algorithm, where we analyzed training loss, validation
performance, and error metrics such as Mean Squared Error
(MSE) and Mean Absolute Error (MAE). Results suggest that
the teleoperation system performs better in basic tasks, whereas
the egocentric condition performed slightly better in complex
tasks. The behavior cloning algorithm demonstrated that the
teleoperation system had stronger generalization across all tasks
compared to the egocentric system.
Submission Number: 6
Loading