by F Stulp
Abstract:
In motor control, high-level goals must be expressed in terms of low-level motor commands. An effective approach to bridge this gap, widespread in both nature and robotics, is to acquire a set of temporally extended actions, each designed for specific goals and task contexts. An action selection module then selects the appropriate action in a given situation. In this approach, high-level goals are mapped to actions, and actions produce streams of motor commands. The first mapping is often ambiguous, as several actions or action parameterizations can achieve the same goal. Instead of choosing an arbitrary action or parameterization, the robot should select those that best fulfill some pre-specified requirement, such as minimal execution duration, successful execution, or coordination of actions with others. The key to being able to perform this selection lies in prediction. By predicting the performance of different actions and action parameterizations, the robot can also predict which of them best meets the requirement. Action models, which have many similarities with human forward models, enable robots to make such predictions. In this dissertation, we introduce a computational model for the acquisition and application of action models. Robots first learn action models from observed experience, and then use them to optimize their performance with the following methods: 1) \textbackslashemphSubgoal refinement, which enables robots to optimize actions in action sequences by predicting which action parameterization leads to the best performance. 2) \textbackslashemphCondition refinement and \textbackslashemphsubgoal assertion, with which robots can adapt existing actions to novel task contexts and goals by predicting when action execution will fail. 3) \textbackslashemphImplicit coordination, in which multiple robots globally coordinate their actions, by locally making predictions about the performance of other robots. The acquisition and applications of action models have been realized and empirically evaluated in three robotic domains: the \textbackslashpioneer robots of our RoboCup mid-size league team, a simulated B21 in a kitchen environment, and a PowerCube robot arm. The main principle behind this approach is that in robot controller design, knowledge that robots learn themselves from observed experience complements well the abstract knowledge that humans specify.
Reference:
Tailoring Robot Actions to Task Contexts using Action Models (F Stulp), PhD thesis, Technische Universität München, 2007.
Bibtex Entry:
@phdthesis{stulp_tailoring_2007,
author = {F Stulp},
title = {Tailoring Robot Actions to Task Contexts using Action Models},
school = {Technische Universität München},
year = {2007},
abstract = {In motor control, high-level goals must be expressed in terms of low-level
motor commands. An effective approach to bridge this gap, widespread
in both nature and robotics, is to acquire a set of temporally extended
actions, each designed for specific goals and task contexts. An action
selection module then selects the appropriate action in a given situation.
In this approach, high-level goals are mapped to actions, and actions
produce streams of motor commands. The first mapping is often ambiguous,
as several actions or action parameterizations can achieve the same
goal. Instead of choosing an arbitrary action or parameterization,
the robot should select those that best fulfill some pre-specified
requirement, such as minimal execution duration, successful execution,
or coordination of actions with others. The key to being able to
perform this selection lies in prediction. By predicting the performance
of different actions and action parameterizations, the robot can
also predict which of them best meets the requirement. Action models,
which have many similarities with human forward models, enable robots
to make such predictions. In this dissertation, we introduce a computational
model for the acquisition and application of action models. Robots
first learn action models from observed experience, and then use
them to optimize their performance with the following methods: 1)
{\textbackslash}{emphSubgoal} refinement, which enables robots to
optimize actions in action sequences by predicting which action parameterization
leads to the best performance. 2) {\textbackslash}{emphCondition}
refinement and {\textbackslash}emphsubgoal assertion, with which
robots can adapt existing actions to novel task contexts and goals
by predicting when action execution will fail. 3) {\textbackslash}{emphImplicit}
coordination, in which multiple robots globally coordinate their
actions, by locally making predictions about the performance of other
robots. The acquisition and applications of action models have been
realized and empirically evaluated in three robotic domains: the
{\textbackslash}pioneer robots of our {RoboCup} mid-size league team,
a simulated B21 in a kitchen environment, and a {PowerCube} robot
arm. The main principle behind this approach is that in robot controller
design, knowledge that robots learn themselves from observed experience
complements well the abstract knowledge that humans specify.},
url = {http://mediatum2.ub.tum.de/node?id=617105},
}