Training Robots without Robots: Deep Imitation Learning for Master-to-Robot Policy Transfer

TL;DR

This work tackles the challenge of transferring force-feedback manipulation skills to robots without requiring robot-based demonstrations. It proposes a Master-to-Robot (M2R) policy transfer system that leverages a master controller with identical kinematics, gaze-based visual attention to bridge visual gaps, dual-action networks for broad-to-precise control, and Transformer-based attention over force/torque signals to infer policies. The approach is validated on a bottle-cap-opening task, achieving an 83.3% final success rate and demonstrating the effectiveness of master-only demonstrations, kinematic calibration, and perceptual attention in enabling force-feedback manipulation. The proposed method reduces cost and safety concerns, enabling scalable data collection and transfer to robots, with potential applicability to other force-sensitive manipulation tasks.

Abstract

Deep imitation learning is promising for robot manipulation because it only requires demonstration samples. In this study, deep imitation learning is applied to tasks that require force feedback. However, existing demonstration methods have deficiencies; bilateral teleoperation requires a complex control scheme and is expensive, and kinesthetic teaching suffers from visual distractions from human intervention. This research proposes a new master-to-robot (M2R) policy transfer system that does not require robots for teaching force feedback-based manipulation tasks. The human directly demonstrates a task using a controller. This controller resembles the kinematic parameters of the robot arm and uses the same end-effector with force/torque (F/T) sensors to measure the force feedback. Using this controller, the operator can feel force feedback without a bilateral system. The proposed method can overcome domain gaps between the master and robot using gaze-based imitation learning and a simple calibration method. Furthermore, a Transformer is applied to infer policy from F/T sensory input. The proposed system was evaluated on a bottle-cap-opening task that requires force feedback.

Figures (10)

• Figure 1: (a) Master-to-robot (M2R) policy transfer system can teach robots tasks that require force feedback from master-only-generated demonstrations. This method uses gaze-based visual attention to minimize the visual gap between the master and robot. (b) The dual-action kim2021gaze, which separates the entire robot trajectory into the global-action of approximate reaching to the target object and local-action of precise manipulation of the object, is used for the precise reaching subtask. (c) The output foveated vision of the proposed gaze-based visual attention can effectively eliminates the visual gap between the master and the robot.
• Figure 2: Hardware components of the M2R policy transfer system. The master controller (\ref{['fig:gp_master']}) and robot (\ref{['fig:gp_robot']}) share the same DH parameter and the same type of fingertip. (\ref{['fig:camera']}) Camera mount is attached to both the master and robot.
• Figure 3: Example of the successful behavior of the proposed DA-force. DA-force successfully grasped the bottle ($\sim 7s$), grasped the cap ($7s \sim 28s$), and rotated it ($28s \sim 33s$).
• Figure 4: Network architectures; the number in brackets indicates the dimensions of the data.
• Figure 5: Calibration result.