Low-cost wheeled-humanoid robot teleoperation system for imitation learning

Purpose – Robot demonstration data is of great significance for enhancing robot autonomy through data-driven methods such as imitation learning. High-quality robot data is mainly obtained through teleoperation. However, current teleoperation methods either have high costs, poor usability or insufficient support for wheeled humanoid robots. This paper aims to design a low-cost wheeled-humanoid robot teleoperation system to address the above problems. Design/methodology/approach – Our teleoperation system only uses one camera and one microphone to ensure low cost. It includes a fiducial-marker-based dual-arm manipulation module for the teleoperation of robot arms, a human-pose-based locomotion piloting module for the teleoperation of the robot locomotion, a voice command control module for gripper control and mode switching and an error-enhanced imitation learning module to train robot’s policy. Findings – A robot teleoperation system with extremely low cost is constructed, which combines vision and voice to enhance usability. Sufficient experiments show that, in various tasks, with just one operator, the proposed system can efficiently teleoperate the wheeled-humanoid robot, collect high-quality data and train imitation learning policies. Originality/value – A low-cost vision-voice teleoperation system for wheeled-humanoid robots is proposed, which is of great significance for the scalable demonstration data collection and robot imitation learning research.