Sensorimotor Primitives for Robot Skills
(last updated 2/95)
Skills are encapsulated task solutions for robustly executing complex,
recurring tasks in a domain. Integrating skills with high-level plans is an
instantiation of the general principle of combining deliberative planning and
reaction which is widely accepted in robotics today. Devel oping robot skills
would significantly impact the robot programming problem, but skill synthesis
remains an open and difficult problem. I propose a complete system for skill
synthesis based on the cognitive and associative phases of skill
acquisition. The system has three components: 1) sensorimotor primitives
(SMP's), 2) skill programming interface (SPI), and 3) skill tuning. Grounding
the approach are sensorimotor primitives which are sensor-integrated,
task-relevant commands. The idea is to bridge the task space with the
robot/sensor space so that task strategy translation onto the robot/sensor
system is more direct. In addition, integrating sensors for task domains
instead of individual tasks leverages previous work in sensor application to
similar tasks. The SPI provides a graphical programming environment for
combining SMP's into skills in the cognitive phase of skill synthesis. The idea
is to leverage the cognitive capabilities of the human to configure
task-relevant commands into skills while hiding implementation programming
complexity. Finally, the skill tuning module provides the associative phase of
skill synthesis: improving task performance with practice. This capability
complements the SPI module by pro viding a mechanism to fine-tune the
approximate parameter values entered during the cognitive stage. Since the task
representation used during the cognitive phase will probably not correctly
model all the fine details of the actual task, the skill tuning module provides
a mechanism for adapting these parameters based on actual task performance.
jmorrow@ri.cmu.edu