Collocated with ICAPS 2019 in Berkeley, USA.
Automated planners are increasingly being integrated into online execution systems. The integration may, for example, embed a domain-independent temporal planner in a manufacturing system (e.g., the Xerox printer application) or autonomous vehicles. The integration may resemble something more like a "planning stack" where an automated planner produces an activity or task plan that is further refined before being executed by a reactive controller (e.g., robotics). Or, the integration may be a domain-specific policy that maps states to actions (e.g., reinforcement learning). Online learning may or may not be involved, and may include adjusting or augmenting the model, determining when to repair versus replan, learning to switch policies, etc. A specific focus of these integrations involves online deliberation, bringing to the foreground concerns over how much computational effort planning should invest over time. But reality rarely proceeds according to the plan or the model. Planning, plan execution, diagnosis, and causal explanation have each been examined by various research efforts, but discussion of the linkages between them in the literature is still somewhat sparse. When considering how to integrate these functions, at least three questions must be considered: (1) System integration: how to integrate planning, plan execution, diagnosis, and causal explanation in a single system? (2) Model / Belief updates: when the unexpected happens, how does the system change its internal representation so future plans are effective? (3) Replanning: what to do now that the unexpected has happened?
|Welcome and Introduction|
A Hybrid Planning and Execution Approach Through HTN and MCTS
Xenija Neufeld, Sanaz Mostaghim and Diego Perez-Liebana
Automated Verification of Social Laws Robustness for Reactive Agents
Alexander Tuisov and Erez Karpas
Monitoring Numeric Expectations in Goal Reasoing Agents
Noah Reifsnyder and Hector Munoz-Avila
|1030|| Morning Break
|1100|| Session 2
Enabling Limited Resource-Bounded Disjunction in Scheduling
Jagriti Agrawal, Wayne Chi, Steve Chien, Gregg Rabideau, Stephen Kuhn and Daniel Gaines
Interleaving Acting and Planning Using Operational Models
Sunandita Patra, Malik Ghallab, Dana Nau and Paolo Traverso
Executing Multi-Goal Mission Plans for Coordinated Mobile Robots
Marlyse Reeves, Enrique Fernandez Gonzalez and Brian Williams
|1400|| Guest Speaker: Christian Fritz
|Co-located with the Actions Workshop!
Title: Planning in Industry, An Experience Report
Abstract: During my 8+ years working in industry I have encountered several problems that lent themselves to solutions that relied on knowledge representation and planning. This includes automated process planning and designing fixtures for CNC machining, mixed-mode transportation planning, high-level task planning for a mobile service robot, and automated testing of robot-behavior. In this talk I will describe these problems and discuss our solution to a couple of them in detail, focusing on aspects that put some of the assumptions made in academic planning research into question. My objective with this is to convince primarily the students in the audience that in these problems (a) representation, not planning, was the most impactful ingredient to success, that (b) in any automation problem the plan does not matter but execution does, and that (c), unsurprisingly, domain-independence does not matter when faced with problems in any specific domain.
Bio: Christian's background is in knowledge representation and planning. He is currently the Vice President of Software Engineering at Savioke, maker of the Relay robot, an autonomous delivery robot for crowded indoor environments like hotels and hospitals. Prior to joining Savioke Christian led the Representation and Planning Area at PARC, a Xerox Company. Christian earned his BS and MS from RWTH Aachen University, Germany, and his PhD from the University of Toronto, Canada.
|1500|| Poster Session
|1530|| Afternoon break
On Expected Value Strong Controllability
Dynamic Controllability with Single and Multiple Indirect Observations
Paul Morris and Arthur Bit-Monnot
Executing Contingent Plans: Addressing Challenges in Deploying Artificial Agents
Christian Muise, Miroslav Vodolan, Shubham Agarwal, Ondrej Bajgar, Luis Lastras and Josef Ondrej