Dynamic Movement Primitives
Overview
A dynamic movement primitive (DMP) is derived from a non-linear
spring-damper system with gains selected to render it critically
damped. It is perturbed with additional forcing terms that generate
the novel behavior captured in a task demonstration. To do this,
the forcing term must be learned from the demonstration represented
as a trajectory. For example, a robot could be kinesthetically
moved to demonstrate a task as the joint positions are recorded
throughout the demonstration. The forcing term is estimated from
the demonstration and then learned by performing locally-weighted
regression. The major advantage of DMPs is their generalizability.
Intuitively, executing a basic motor behavior is localized in space
and should be flexible in velocity, along with where the motion
starts and ends, while maintaining the fundamental behavior. DMPs
possess this property as both the start and goal locations can be
adjusted and a separate temporal scalar can be used to adapt the
speed of execution. Another important characteristic of DMPs is
their ability to perform online adaptation, i.e., a robot can react
to environmental changes mid-execution. This includes obstacle
avoidance and the adaptation based on force feedback.
Contributions
- We utilized DMPs to represent lunar excavator trajectories enabling us to develop a compact, generalized representation of the robot's motion plans