A robot learning from demonstration framework for skillful small parts assembly

Abstract

Increasing demand for higher production flexibility and smaller production batch size pushes the development of manufacturing expertise towards robotic solutions with fast setup and reprogram capability. Aiming to facilitate assembly lines with robots, the learning from demonstration (LfD) paradigm has attracted attention. A robot LfD framework designed for skillful small parts assembly applications is developed, which takes position, orientation and wrench demonstration data into consideration while utilizes impedance control to deal with the motion error. In view of constraints in industrial assembly applications, we propose a robot LfD framework where policy learning is carried out with separated assembly demonstration data to avoid potential under-fitting problem. With the proposed assembly policies, reference orientation and wrench trajectories are generated as well as coupled with the position data to boost their generalization and robust performance. Effectiveness of the proposed LfD framework is validated by a printed circuit board assembly experiment with a 7-DOF torque-controlled robot.

Robot Learning from Demonstration for Assembly with Camera-Supplemented Optical Motion Capture Sensors

“…J (•) ∈ R 6×7 denote the Jacobian matrix [16], and h(•) represents the robot's dynamics that allows for gravity compensation torque, Coriolis force, and friction. The operators log(•), exp(•), * , • denote the logarithmic map, exponential map, product, and conjugate of the unit quaternion [10]. In each control period, the goal pose update law updates the p g,t and p g,t in (2):…”

“…All of those movements can be viewed as the screw movements with finite pitch [16]. The rotation movement primitive consists of the controller (2) and the goal update law (10). p g,t+1 = p g,t q g,t+1 = exp δ t β q d s * q g,t (10) where β q > 0 is the goal orientation update coefficient.…”

“…Hence, assembling those objects is challenging for the robot. The methods in [10,12,29] learn a pose trajectory for each phase (sub-task) from multiple demonstration trajectories. However, regarding the incontact assembly task [14], those pose-tracking based methods suffer from low success rate in the aligning and assembling phase.…”

“…What they do is nothing other than what they used to do in the assembly lines [13]. In contrast to the widely-applied teleoperation [21] and kinesthetic teaching [10] methods, the MoCap-based methods are applicable without the absence of the robots.…”

“…Nevertheless, when the assembly action is rotation, i.e. χj,k = 2, we first map all the orientation q n,j to the tangent space [10] of the first via the quaternion logarithmic operation in each iteration of the DP algorithm. Then the cutting point is determined in the tangent space by the same method as the one for the translation action.…”