Abstract:
In this thesis we study a reasoning module for agents that have cognitive abilities, such as memory, perception, action, and are expected to function autonomously for long periods of time. The module provides the ability to reason about action and change using the language of the situation calculus and variants of the basic action theories. The main focus of this thesis is on the logical problem of progressing an action theory.

First, we investigate the conjecture by Lin and Reiter that a practical first-order
definition of progression is not appropriate for the general case. We show that Lin and Reiter were indeed correct in their intuitions by providing a proof for the conjecture, thus resolving the open question about the first-order definability of progression and justifying the need for a second-order definition.

Then we proceed to identify three cases where it is possible to obtain a first-order progression with the desired properties: i) we extend earlier work by Lin and Reiter and present a case where we restrict our attention to a practical class of queries that may only quantify over situations in a limited way; ii) we revisit the local-effect assumption of Liu and Levesque that requires that the effects of an action are fixed by the arguments of the action and show that in this case a first-order progression is suitable; iii) we investigate a way that the local-effect assumption can be relaxed and show that when the initial knowledge base is a database of possible closures and the effects of the actions are range-restricted then a first-order progression is also suitable under a just-in-time assumption.

Finally, we examine a special case of the action theories with range-restricted effects and present an algorithm for computing a finite progression. We prove the correctness and the complexity of the algorithm, and show its application in a simple example that is inspired by video games.

Bibtex:

@phdthesis{vassos09phdthesis,
address = {Toronto, Canada},
author = {Vassos, Stavros},
school = {University of Toronto},
title = {A Reasoning Module for Long-Lived Cognitive Agents},
year = {2009}
}

Bibtex:

@inproceedings{vassos09rangerestricted,
address = {Toronto, Canada},
author = {Vassos, Stavros and Sardina, Stavros and Levesque, Hector},
booktitle = {Proceedings of the Ninth International Symposium on Logical Formalizations of Commonsense Reasoning (CS-09)},
editor = {Lakemeyer, Gerhard and Morgenstern, Leora and Williams, Mary-Anne},
pages = {135–140},
publisher = {UTSePress},
title = {Progressing basic action theories with non-local effect actions},
year = {2009}
}

Bibtex:

@inproceedings{vassos09localeffect,
address = {Sydney, Australia},
author = {Vassos, Stavros and Lakemeyer, Gerhard and Levesque, Hector},
booktitle = {Proceedings of 11th International Conference on Principles of Knowledge Representation and Reasoning (KR-08)},
editor = {Brewka, Gerhard and Lang, J{\’{e}}r{\^{o}}me},
pages = {662–672},
publisher = {AAAI Press},
title = {First-Order Strong Progression for Local-Effect Basic Action Theories},
year = {2009}
}

This paper was selected for the “Outstanding Paper Honorable Mention Award” in the Twenty-Third AAAI Conference on Artificial Intelligence!
Click here for more details about the award.

Abstract:
In a seminal paper, Lin and Reiter introduced a model-theoretic definition for the progression of the initial knowledge base of a basic action theory. This definition comes with a strong negative result, namely that for certain kinds of action theories, first-order logic is not expressive enough to correctly characterize this form of progression, and second-order axioms are necessary. However, Lin and Reiter also considered an alternative definition for progression which is always first-order definable. They conjectured that this alternative definition is incorrect in the sense that the progressed theory is too weak and may sometimes lose information. This conjecture, and the status of first-order definable progression, has remained open since then. In this paper we present two significant results about this alternative definition of progression. First, we prove the Lin and Reiter conjecture by presenting a case where the progressed theory indeed does lose information. Second, we prove that the alternative definition is nonetheless correct for reasoning about a large class of sentences, including some that quantify over situations. In this case the alternative definition is a preferred option due to its simplicity and the fact that it is always first-order.

Bibtex:

@inproceedings{vassos08conjecture,
author = {Vassos, Stavros and Levesque, Hector},
booktitle = {Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (AAAI-08)},
location = {Chicago, Illinois, USA},
month = {July},
pages = {1004–1009},
title = {On the Progression of Situation Calculus Basic Action Theories: Resolving a 10-year-old Conjecture},
year = {2008}
}

Dear Stavros and Hector,

On behalf of the Association for the Advancement of Artificial Intelligence and the AAAI-08 Program Committee, we are pleased to inform you that your paper, entitled “On the Progression of Situation Calculus Basic Action Theories: Resolving a 10-year-old Conjecture,” has been selected for Honorable Mention at AAAI-08.

Each year AAAI’s Conference on Artificial Intelligence honors papers that exemplify high standards in technical contribution and exposition. Papers were recommended for this status by members of the Program Committee during the blind review process. The winning papers were selected by the Program Chairs with the help of some members of the Senior Program Committee.

In recognition of your achievement, you and your coauthors will be presented with certificates by the AAAI-08 Program Cochairs, Dieter Fox and Carla Gomes, during their opening remarks on Tuesday, July 15, at 8:30 am. There will also be an announcement of this honor in the front matter of the AAAI-08 Proceedings.
[...]

More on the paper here and more on the award here.

Abstract:

In this paper, we propose a new progression mechanism for a restricted form of incomplete knowledge formulated as a basic action theory in the situation calculus. Specifically, we focus on functional fluents and deal directly with the possible values these fluents may have and how these values are affected by both physical and sensing actions. The method we propose is logically complete and can be calculated efficiently using database techniques under certain reasonable assumptions.

Bibtex:

@inproceedings{vassos07progression,
title = {Progression of Situation Calculus Action Theories with Incomplete Information},
author = {Vassos, Stavros and Levesque, Hector },
booktitle = {Proceedings of the 20th International Joint Conference on Artificial Intelligence},
editor = {Veloso, Manuela M. },
pages = {2024–2029},
address = {Hyderabad, India},
citeulike-article-id = {1923630},
keywords = {ai, progression, reasoning_about_action, situation_calculus},
month = {January},
year = {2007}
}

Giving robots the ability to think and reason in an industrial environment
By Mary Del Ciancio

Imagine a manufacturing plant where robots are sophisticated enough to understand their environment and choose the best path to achieve a goal. Imagine robots capable of working together as a team to solve problems. Such robots, endowed with high-level cognitive capabilities – including perception processing, attention allocation, anticipation, planning and reasoning – would greatly expand the possibility of flexible manufacturing because they would be able to see, feel, touch and reason within the confines of unpredictable environments.

This isn’t your father’s robot.

Click here to read the rest of the article in the “Manufacturing Automation” magazine website.

*[part of the text used in the answer to some questions is taken from the webpage of the Cognitive Robotics group at UofT]

1. What is your definition of cognitive robotics?

The term “cognitive robotics” is used to refer to robots with higher level cognitive functions that involve knowledge representation and reasoning. A cognitive robot has an internal representation of its own abilities to act and perceive information as well as the current state of the environment. This allows the robot to reason and synthesize an appropriate course of action for each of the situations it may encounter as a result of reasoning about goals that need to be fulfilled, the available actions that can be performed, sensor data that the robot perceives from the environment, the mental states of other robots, etc.

- What makes a robot a cognitive one? What sets it apart from other robots?

Every robotic system has a kind of core program that specifies what are the actions that the robot should perform at every point in time (e.g. move arm towards object, open gripper, etc). This program is done by a programmer who takes into account the task that the robot should do, the mechanical actions it can perform, the information the robot can get from the environment, etc, and constructs a program that when executed will realize the intended behavior.

In a sense, most of the logic behind the structure of the program is implicit as it only exists in the mind of the programmer. The intended behavior for the robot might be something like “whenever there exists an object in some specific boundaries, the robot should pick it up and move it to the appropriate box”. Even though this might be apparent when inspecting the code, the actual logic behind this is not explicit in the program.

The difference with cognitive robotics is that instead of programming a robot with the sequences of mechanical actions it should perform when certain conditions arise, a cognitive robot is programmed with the logic behind the mechanical actions it can perform and how to synthesize these mechanical actions in order to bring the environment to a specific state. Any task that the robot should do can then be expressed in a more abstract way and the actual sequences of actions that should be performed will be computed by the program itself.

- How is it different from vision-guided robotics?

The notion of cognitive robotics is somewhat orthogonal to vision-guided robotics. The central feature of cognitive robotics is the explicit representation of the interaction between the robot and its environment, and the ability of the robot to reason within this representation so as to specify the actions that should be performed in order to complete a task. As the the robot is reasoning to compute the course of action the sensor data is also taken into account and can be used to guide it. This sensor data can be a set of features extracted by visual cameras, laser or infrared sensors, or any other kind of sensors that are appropriate.

2. What is the goal of cognitive robotics?

Similarly to question 1, cognitive robotics is concerned with endowing robots with higher level cognitive functions that involve reasoning, for example, reasoning about goals, perception, actions, the mental states of other robots, collaborative task execution, etc.

3. What industrial applications are we seeing cognitive robotics applied in today?

To my knowledge there are no industrial applications where the cognitive robotics design paradigm is used.

- What applications will we see in the future?

In the past robots have replaced humans in many tasks in the industry, mainly tasks that involved certain skills and minimal reasoning capabilities. It is possible that as the techniques related to cognitive robotics become more powerful and the implementations of cognitive robots more robust, we will see cognitive robots replace humans in more advanced tasks that currently robots cannot safely perform.

Also, the cognitive robotics design paradigm offers some advantages with respect to modularity and adaptability, even though reasoning may not be necessary for a robot in an industrial application. The fact that the behavior of a cognitive robot is a result of reasoning about the ways it can interact with the environment and the specification of the task it should do, enables it to adapt to different settings. For instance, a futuristic scenario may involve cognitive robots that consist of modules that come with a representation of their parts and the dynamics of the available mechanical actions they can perform. The modules can then be combined and incorporated into industrial applications by programming only at the level of task specification, while the low-level details are sorted out by reasoning that the cognitive robot does by itself.

4. What is needed in order for cognitive robotics to perform? (i.e. software, hardware, etc.)

Cognitive robotics is a design paradigm that when applied to robotic agents (in contrast to software agents) it involves taking care of issues that lie in several different fields of research and application. In particular there are the following parts:

1. The mechanical part of the robot that is responsible for movement and affecting the environment (e.g. mechanical arms, the body, the motors).
2. The software and hardware part responsible for getting meaningful information from the environment that the robot is situated (e.g. the hardware and the information processing software for doing feature extraction from visual images and sound).
3. The software part responsible for the representation of the environment and the way that the robot can interact with it (e.g. a logical specification of the properties of the environment as well as how these are affected by the available actions that the robot can perform).
4. The software part responsible for the specification of the task that the robot should do based on the previous representation.
5. The software and hardware part that makes use of the representation of the environment and the specification of the task that the robot should do (numbers 3 and 4) in order to compute the behavior of the robot at any given moment.
6. The software and hardware part that provides the interface between the reasoning component (number 5) and the actual sensors and actuators in the environment (numbers 1 and 2).

5.What are the challenges associated with cognitive robotics?

Even though each of the parts discussed in the previous question comes with its own set of challenges, the synthesis of those parts into a robust cognitive robotics architecture is probably the biggest challenge. In particular for the case of cognitive robots for industrial applications it is needed that people from the relevant fields of academic research and the industry are brought together in order to identify a specific real-world industrial application that will motivate approaches that deal with the problem as a whole.

6. What robotics “problems” does cognitive robotics fix?

Cognitive robotics is a design paradigm that separates the “high-level” intended behavior of the robot from the “low-level” details of how it can be realized. Depending on the requirements for a robotic application this may be beneficial with respect to adaptability and modularity. Also, the knowledge representation and reasoning part provides a clean framework for specifying advanced tasks related to proactive behavior (e.g. short-term planning before acting in order to find a particular sequence of actions that can achieve a specific goal).

7. What are some of the really amazing applications you’ve seen cognitive robotics applied in?

(Not an industrial application). There is an ongoing project regarding unmanned aerial vehicles that uses fully autonomous helicopters for finding survivors in rescue missions.

8. Tell me about the research your group is doing in cognitive robotics.

Part of the group’s focus is research related to a high-level programming language for controlling robots called IndiGolog. In this language one can specify the following:

• The available actions that the robot can do, including actions that affect the environment and actions that extract information from the environment using its sensors.
• The objects and properties of the environment and how these change for each of the possible actions the robot can do.
• A high-level control program that specifies the behavior of the robot. This looks like an ordinary imperative program but uses the properties of the environment as control-flow expressions and the actions that the robot can do as basic statements. Moreover, the program may include statements expressing that the robot should act accordingly in order to make a certain property of the environment become true. For this kind of statements the interpreter will have to reason at run-time in order to compute a sequence of actions that will realize the intended property based on the conditions that hold at run-time.

A prototype implementation of the language has been developed and experiments have been conducted using the language to build a high-level controller for simple robots such as the Sony ERS7 Aibo.

- Do you partner with industry/manufacturers for research? Find out what they’d like robots to do?

There is no close connection with the industry or manufacturers at the moment. As mentioned also in question 5, it is one of the current challenges to bring the two communities closer together.

Abstract:

This paper describes an implementation of the the Wumpus World in IndiGolog with the objective of showing the applicability of this interleaved agent programming language for modeling agent behavior in realistic domains. We briefly go over the IndiGolog architecture, explain how we can reason about the Wumpus World domain, and show how to express agent behavior using high-level agent programs. Finally, we discuss initial empirical results obtained as well as challenging issues to be resolved.

Bibtex:

@inproceedings{sardina05wumpus,
title = {The {W}umpus {W}orld in {I}ndi{G}olog: {A} {P}reliminary {R}eport},
author = {Sardina, Sebastian and Vassos, Stavros },
booktitle = {In Proceedings the Nonmonotonic Reasoning,
Action and Change Workshop at IJCAI (NRAC-05)},
address = {Edinburgh, Scotland},
pages = {90–95},
citeulike-article-id = {381288},
keywords = {agent, ai, cogrobo, games, golog, incomplete_knowledge, logic, reasoning_about_action, situation_calculus, system, wumpus},
year = {2005}
}

Abstract:
In this thesis we present Lp, a reinterpretation of situation calculus based on intuitions from many-valued logics. The key difference is that the notion of truth is based on the fact that a term is interpreted into a set of objects rather than one single object and equality is interpreted as “possibly equals”. Lp is suitable for defining action theories that capture fluent-based disjunctive knowledge, which means that any incomplete knowledge the theory captures is limited to be about the value of one fluent each time. This essentially enforces an independence assumption on the fluents which allows for efficient evaluation mechanisms. We show that like situation calculus a similar regression theorem holds in Lp. Furthermore, we prove that Lp can be embedded in situation calculus and show how a special form of Lp theories can be soundly implemented in Prolog and the agent programming language Indigolog.

Bibtex:

@mastersthesis{vassos05msthesis,
title = {A Feasible Approach to Disjunctive Knowledge in Situation Calculus},
author = {Vassos, Stavros },
editor = {Levesque, Hector and Bacchus, Fahiem },
school = {University of Toronto},
citeulike-article-id = {2032867},
keywords = {ai, incomplete_knowledge, reasoning_about_action, situation_calculus},
year = {2005}
}