| Summary: | Robot learning from demonstration is a powerful approach which allows the robot to acquire training examples from human demonstrations. These demonstration examples can be obtained in di erent ways, such as recording state-action pairs whilst the robot is tele-operated by a human supervisor or recording the data as the robot observes the human teacher executing the desired behaviour. The tele-operation approach provides the most direct method for information transfer, but what is missing is an active and bidirectional interface in which both actors, the human teacher and the learner robot, can interact in order to improve the overall system's performance during the execution of a given task. The goal of this dissertation is the development of a haptic interface for the tele-operation of the lower limbs of a real humanoid robot during postural balance tasks. At the core of this co-adaptive design is the use of a Phantom Omni haptic device that provides the human teacher with information about the robot's state (e.g., robot's balance), its physical capabilities and/or limitations. This information from the environment is translated in force feedback to allow the user to guide and correct the executed behaviour. At the end of the demonstration phase, data logging of all the sensory information available (e.g, joint positions/velocities, ground reaction forces, inertial data) and the control commands supplied by the human operator are carried out by a suitable software architecture based on ROS (Robot Operating System). These recorded datasets may be useful in future works for implementing a learning algorithm, enabling the robot to reproduce the task without assistance.
|
|---|