An incremental approach to learning generalizable robot tasks from human demonstration

Abstract

DMPs are a common method for learning a control policy for a task from demonstration.\ud This control policy consists of differential equations that can create a smooth trajectory to a new goal point.\ud However, DMPs only have a limited ability to generalize the demonstration to new environments and solve problems such as obstacle avoidance.\ud Moreover, standard DMP learning does not cope with the noise inherent to human demonstrations.\ud Here, we propose an approach for robot learning from demonstration that can generalize noisy task demonstrations to a new goal point and to an environment with obstacles.\ud This strategy for robot learning from demonstration results in a control policy that incorporates different types of learning from demonstration,\ud which correspond to different types of observational learning as outlined in developmental psychology

Robot learning from demonstrations: Emulation learning in environments with moving obstacles

“…In fact, it helps the convergence very much and has negligible influence on the computed optimal parameters. Consequently, we experienced a convergence behaviour much better than Ghalamzan et al (2015) by adding L ∞ of the error signal in eq. (4).…”

“…By contrast, Ghalamzan et al (2015) presented a modular approach, namely IRLfD, that results in local modification of the trajectory for avoiding an obstacle. This method includes: (i) regression, (ii) DMP and (iii) obstacle avoidance modules.…”

“…We use a numerical optimisation approach (minConf ) with a quasi-Newton strategy and limited-memory BFGS updates Schmidt (2010). Ghalamzan et al (2015) proposed to use only L 2 norm of the error signal. This has a very small and even a zero gradient in some parts of the parameter space θ .…”

“…Similar to proposals in studies of sensorimotor learning in cognitive psychology and neuroscience (Todorov, 2004;Wolpert et al, 2011), we design the reproduction phase to cope with the moving obstacles. The approach proposed by Ghalamzan et al (2015) is an open loop strategy in which the position of the obstacle is assumed to be known before generating a collision-free trajectory and it does not change during the whole experiment. By contrast, we postulate that a skill, which is learnt via a utility function for a static scene, can be generalised to a dynamic environment by extending the UF to include a feedforward and a feedback strategy.…”

Robot learning of industrial assembly task via human demonstrations

“…GMM can also be combined with virtual spring-damper system (DS-GMR) (Calinon et al 2014) or with linear attractor system combined with GMR (Mühlig et al 2012) so to enable the robot to perform the learned skill with new start and target positions. Moreover, GMM/GMR is combined with DMP for generalizing the learned trajectory to new goal point and obstacle avoidance (Ghalamzan et al 2015).…”