Wed., June 11, 9:00 - 10:00 | |
Speaker:
|
Subbarao Kambhampati (Arizona State University)
|
Chair:
|
Enrico Giunchiglia (University of Genova)
|
Title:
|
1001 Ways to Skin a Planning Graph for Heuristic
Fun and Profit
|
Abstract:
|
I will discuss our experiences in extracting heuristics from planning
graphs to control search in different planners (state-space,
plan-space, disjunctive) and for different planning problems
(classical, metric-temporal, conformant and conditional).
You can find the slides of the presentation here. |
Thu., June 12, 9:00 - 10:00 | |
Speaker:
|
Malik Ghallab (LAAS-CNRS)
|
Chair:
|
Nicola Muscettola (Nasa AMES Research Center)
|
Title:
|
Plan-Based Robot Control
|
Abstract:
|
Automated planning techniques are certainly essential for the design
of systems exhibiting autonomous deliberative behaviors. But these
techniques do not seem today to be a critical component of robotics
research. This is probably due to the state of the art in both,
planning and robotics, to the demanding requirements of robotics
applications for handling uncertainty, highly unstructured domains,
and less predictable actions, and to the restricted use in robotics
in which planning techniques have been focused.
Planning techniques are already mature enough for the synthesis of
high level mission plans with correct management of temporal and
resources constraints. But the "elementary actions" of such plans are
at an abstract level, far away from the low-level sensory-motor
primitives of a robot. However, planning techniques can also be very
useful for the robust decomposition and control of planned actions
into execution primitives.
The talk introduces briefly the mission level planning in robotics
and its relationship with motion planning to take into account space
constraints and the kinematics of the robot. It then focuses on the
robust execution of such a plan, that is how to decompose an abstract
action such as "goto(position)", or "extract-sample(from)" into a
complex closed-loop control of sensory-motor primitives.
The approach is to specify a collection of Hierarchical Tasks
Networks (HTN), called modalities, whose primitives are sensory-motor
functions. Each modality is a possible way of combining some of these
low-level functions to achieve the desired task. The relationship
between control states and the appropriate modality for pursuing a
task is learned through experience as a Markov Decision Process
(MDP). This MDP automata provides a general policy for the task. It
is independent of a particular environment. It characterizes the
robot abilities for that task. It can be learned off-line, as part of
the robot design and programming.
The approach has been extensively experimented with and characterized
for an indoor robot. It is also being deployed on an outdoor platform.
|
Fri., June 13, 9:00 - 10:00 | |
Speaker:
|
Douglas Smith (Kestrel Institute)
|
Chair:
|
Dana Nau (University of Maryland)
|
Title:
|
Automated Synthesis of High-Performance Planners and Schedulers
|
Abstract:
|
Program synthesis is the automated generation of code from
formal specifications. We have explored a variety of approaches to
applying synthesis technology to the production of high-performance
planners and schedulers. Our current thinking is embodied in a
generator called Planware II which uses a domain-specific
specification formalism based on state machines to model complex
resource systems. The state machines, better thought of as activity
machines, include constraints on activities and their transitions, as
well as services that model the interaction between resources.
Program schemas are instantiated to yield fast, customized search and
propagation code. This talk will survey our technological progress on
synthesizing scheduling algorithms and the various applications that
we have developed.
|