Learning to sequence movement primitives from demonstrations

Abstract

Abstract-We present an approach for learning sequential robot skills through kinesthetic teaching. The demonstrations are represented by a sequence graph. Finding the transitions between consecutive basic movements is treated as classification problem where both Support Vector Machines and Gaussian Mixture Models are evaluated as classifiers. We show how the observed primitive order of all demonstrations can help to improve the movement reproduction by restricting the classification outcome to the currently executed primitive and its possible successors in the graph. The approach is validated with an experiment in which a 7-DOF Barrett WAM robot learns to unscrew a light bulb.

Learning movement synchronization in multi-component robotic systems

“…The prevalent technique to model complex tasks in PbD is to use a two-level hierarchy where a top-level discrete symbolic layer controls the switching behavior of underlying action/movement primitives [2], [3], [4], [5]. In these approaches, the symbolic level is typically modeled using directed graphs or finite state machines.…”

“…In these approaches, the symbolic level is typically modeled using directed graphs or finite state machines. For example, in [5] kinesthetic demonstrations are used to teach a robotic hand and arm to unscrew a light bulb. There, learning the skill consists of constructing an FSM where the states are primitives, and transitions signify switching between primitives, of which only one can be active at a time.…”

“…The algorithm used in this stage is similar in structure to that used in [5]. The algorithm starts by searching the entire TPN for repetitions.…”

Probabilistic progress prediction and sequencing of concurrent movement primitives

“…In this paper, each graph contains one node for each activated MP and one transition for each observed pair of succeeding MPs. For more sophisticated sequential graph structures, the interested reader may have a look at our previous work [17]. …”

“…Depending on the control variable, a goal can be a desired position or orientation of a robot body, joint angle, force or a combination thereof and can be defined relative between bodies using reference frames. In previous work, we showed how the attractor goals and reference frames can be extracted from kinesthetic demonstrations [16], [17]. For the remainder of the paper, we therefore impose the following assumptions: We assume to know the number of concurrent sequences.…”