Programming by demonstration (PbD) is an intuitive approach to impart a task to a robot from one or several demonstrations by the human teacher. The acquisition of the demonstrations involves the solution of the correspondence problem when the teacher and the learner differ in sensing and actuation. Kinesthetic guidance is widely used to perform demonstrations. With such a method, the robot is manipulated by the teacher and the demonstrations are recorded by the robot’s encoders. In this way, the correspondence problem is trivial but the teacher dexterity is afflicted which may impact the PbD process. Methods that are more practical for the teacher usually require the identification of some mappings to solve the correspondence problem. The demonstration acquisition method is based on a compromise between the difficulty of identifying these mappings, the level of accuracy of the recorded elements and the user-friendliness and convenience for the teacher. This thesis proposes an inertial human motion tracking method based on inertial measurement units (IMUs) for PbD for pick-and-place tasks. Compared to kinesthetic guidance, IMUs are convenient and easy to use but can present a limited accuracy. Their potential for PbD applications is investigated.

To estimate the trajectory of the teacher’s hand, 3 IMUs are placed on her/his arm segments (arm, forearm and hand) to estimate their orientations. A specific method is proposed to partially compensate the well-known drift of the sensor orientation estimation around the gravity direction by exploiting the particular configuration of the demonstration. This method, called heading reset, is based on the assumption that the sensor passes through its original heading with stationary phases several times during the demonstration. The heading reset is implemented in an integration and vector observation algorithm. Several experiments illustrate the advantages of this heading reset.

A comprehensive inertial human hand motion tracking (IHMT) method for PbD is then developed. It includes an initialization procedure to estimate the orientation of each sensor with respect to the human arm segment and the initial orientation of the sensor with respect to the teacher attached frame. The procedure involves a rotation and a static position of the extended arm. The measurement system is thus robust with respect to the positioning of the sensors on the segments. A procedure for estimating the position of the human teacher relative to the robot and a calibration procedure for the parameters of the method are also proposed. At the end, the error of the human hand trajectory is measured experimentally and is found in an interval between $28.5$ mm and $61.8$ mm. The mappings to solve the correspondence problem are identified. Unfortunately, the observed level of accuracy of this IHMT method is not sufficient for a PbD process.

In order to reach the necessary level of accuracy, a method is proposed to correct the hand trajectory obtained by IHMT using vision data. A vision system presents a certain complementarity with inertial sensors. For the sake of simplicity and robustness, the vision system only tracks the objects but not the teacher. The correction is based on so-called Positions Of Interest (POIs) and involves 3 steps: the identification of the POIs in the inertial and vision data, the pairing of the hand POIs to objects POIs that correspond to the same action in the task, and finally, the correction of the hand trajectory based on the pairs of POIs. The complete method for demonstration acquisition is experimentally evaluated in a full PbD process. This experiment reveals the advantages of the proposed method over kinesthesy in the context of this work.