The judicious use of abstraction can help planning agents to identify key interactions between actions, and resolve them, without getting bogged down in details. However, ignoring the wrong details can lead agents into building plans that do not work, or into costly backtracking and replanning once overlooked interdependencies come to light. We claim that associating systematically-generated summary information with plans' abstract operators can ensure plan correctness, even for asynchronously-executed plans that must be coordinated across multiple agents, while still achieving valuable efficiency gains. In this paper, we formally characterize hierarchical plans whose actions have temporal extent, and describe a principled method for deriving summarized state and metric resource information for such actions. We provide sound and complete algorithms, along with heuristics, to exploit summary information during hierarchical refinement planning and plan coordination. Our analyses and experiments show that, under clearcut and reasonable conditions, using summary information can speed planning as much as doubly exponentially even for plans involving interacting subproblems.

To construct a perpetual self-aware cognitive agent that can continuously operate with independence, an introspective machine must be produced. To assemble such an agent, it is necessary to perform a full integration of cognition (planning, understanding, and learning) and metacognition (control and monitoring of cognition) with intelligent behaviors. The failure to do this completely is why similar, more limited efforts have not succeeded in the past. I outline some key computational requirements of metacognition by describing a multi- strategy learning system called Meta-AQUA and then discuss an integration of Meta-AQUA with a nonlinear state-space planning agent. I show how the resultant system, INTRO, can independently generate its own goals, and I relate this work to the general issue of self-awareness by machine.

The primary revolution in automated planning in the last decade has been the very impressive scale-up in planner performance. A large part of the credit for this can be attributed squarely to the invention and deployment of powerful reachability heuristics. Most, if not all, modern reachability heuristics are based on a remarkably extensible data structure called the planning graph, which made its debut as a bit player in the success of GraphPlan, but quickly grew in prominence to occupy the center stage. Planning graphs are a cheap means to obtain informative look-ahead heuristics for search and have become ubiquitous in state-of-the-art heuristic search planners. We present the foundations of planning graph heuristics in classical planning and explain how their flexibility lets them adapt to more expressive scenarios that consider action costs, goal utility, numeric resources, time, and uncertainty.

Most classical scheduling formulations assume a fixed and known duration for each activity. In this paper, we weaken this assumption, requiring instead that each duration can be represented by an independent random variable with a known mean and variance. The best solutions are ones which have a high probability of achieving a good makespan. We first create a theoretical framework, formally showing how Monte Carlo simulation can be combined with deterministic scheduling algorithms to solve this problem. We propose an associated deterministic scheduling problem whose solution is proved, under certain conditions, to be a lower bound for the probabilistic problem. We then propose and investigate a number of techniques for solving such problems based on combinations of Monte Carlo simulation, solutions to the associated deterministic problem, and either constraint programming or tabu search. Our empirical results demonstrate that a combination of the use of the associated deterministic problem and Monte Carlo simulation results in algorithms that scale best both in terms of problem size and uncertainty. Further experiments point to the correlation between the quality of the deterministic solution and the quality of the probabilistic solution as a major factor responsible for this success.

This paper describes Marvin, a planner that competed in the Fourth International Planning Competition (IPC 4). Marvin uses action-sequence-memoisation techniques to generate macro-actions, which are then used during search for a solution plan. We provide an overview of its architecture and search behaviour, detailing the algorithms used. We also empirically demonstrate the effectiveness of its features in various planning domains; in particular, the effects on performance due to the use of macro-actions, the novel features of its search behaviour, and the native support of ADL and Derived Predicates.

We describe a vision-based obstacle avoidance system for off-road mobile robots. The system is trained from end to end to map raw input images to steering angles. It is trained in supervised mode to predict the steering angles provided by a human driver during training runs collected in a wide variety of terrains, weather conditions, lighting conditions, and obstacle types. The robot is a 50cm off-road truck, with two forwardpointing wireless color cameras. A remote computer processes the video and controls the robot via radio. The learning system is a large 6-layer convolutional network whose input is a single left/right pair of unprocessed low-resolution images. The robot exhibits an excellent ability to detect obstacles and navigate around them in real time at speeds of 2 m/s.

Probabilistic temporal planning attempts to find good policies for acting in domains with concurrent durative tasks, multiple uncertain outcomes, and limited resources. These domains are typically modelled as Markov decision problems and solved using dynamic programming methods. This paper demonstrates the application of reinforcement learning -- in the form of a policy-gradient method -- to these domains. Our emphasis is large domains that are infeasible for dynamic programming. Our approach is to construct simple policies, or agents, for each planning task. The result is a general probabilistic temporal planner, named the Factored Policy-Gradient Planner (FPG-Planner), which can handle hundreds of tasks, optimising for probability of success, duration, and resource use.

Probabilistic temporal planning attempts to find good policies for acting in domains with concurrent durative tasks, multiple uncertain outcomes, and limited resources. These domains are typically modelled as Markov decision problems and solved using dynamic programming methods. This paper demonstrates the application of reinforcement learning -- in the form of a policy-gradient method -- to these domains. Our emphasis is large domains that are infeasible for dynamic programming. Our approach is to construct simple policies, or agents, for each planning task. The result is a general probabilistic temporal planner, named the Factored Policy-Gradient Planner (FPG-Planner), which can handle hundreds of tasks, optimising for probability of success, duration, and resource use.

We describe a vision-based obstacle avoidance system for off-road mobile robots. The system is trained from end to end to map raw input images to steering angles. It is trained in supervised mode to predict the steering angles provided by a human driver during training runs collected in a wide variety of terrains, weather conditions, lighting conditions, and obstacle types. The robot is a 50cm off-road truck, with two forwardpointing wireless color cameras. A remote computer processes the video and controls the robot via radio. The learning system is a large 6-layer convolutional network whose input is a single left/right pair of unprocessed low-resolution images. The robot exhibits an excellent ability to detect obstacles and navigate around them in real time at speeds of 2 m/s.

Probabilistic temporal planning attempts to find good policies for acting in domains with concurrent durative tasks, multiple uncertain outcomes, and limited resources. These domains are typically modelled as Markov decision problems and solved using dynamic programming methods. This paper demonstrates the application of reinforcement learning -- in the form of a policy-gradient method -- to these domains. Our emphasis is large domains that are infeasible for dynamic programming. Our approach isto construct simple policies, or agents, for each planning task. The result is a general probabilistic temporal planner, named the Factored Policy-Gradient Planner (FPG-Planner), which can handle hundreds of tasks, optimising for probability of success, duration, and resource use.