Probabilistic action prospection based on experiences - representation, learning and reasoning in autonomous robotic agents

Abstract

The concept of autonomous robotic companions assisting with tedious tasks or daily routines has long been a futuristic ambition, especially in scenarios deemed too complex for seamless operation. The multitude of variables, intricate interconnections, and numerous (side) effects on seemingly straightforward influences pose significant challenges for them to function competently without human intervention. However, if robots were equipped with knowledge about themselves, their surroundings, and the objects within it, they could address various queries about their environment and undertake tasks independently, drawing further insights from their own experiences and sensory inputs. For example, they could effortlessly respond to contextual inquiries such as “Where should I position myself in the kitchen to locate the milk carton?”, “What route should I take from my current location?” and “Is the refrigerator open or closed?”. Addressing uncertainty and its associated limitations is crucial for constructing a comprehensive world model that provides an autonomous agent with the necessary capabilities to operate independently – potentially leading to robots becoming valuable household aids and companions. This thesis presents BayRoB, a probabilistic framework integrating probabilistic hybrid action models to assist autonomous agents in making informed, context-driven decisions under uncertainty. BayRoB utilizes probabilistic models to represent an autonomous robot’s belief state and offers mechanisms to track changes in this state over time. The framework incorporates a novel formalism that enables the learning, representation of and reasoning over joint probability distributions representing action and object designators. Enabling robots to adeptly handle uncertain situations significantly enhances their decision-making abilities and contributes to their capacity to anticipate action outcomes and environmental changes, thereby promoting autonomy. The approach of integrating probabilistic hybrid models into a framework, as demonstrated by BayRoB, with learning occurring through experiential data, holds promise for fundamentally enhancing the decision-making processes of autonomous agents. A critical aspect of this advancement lies in the incorporation of joint probability distributions, encompassing both aspects of the world and the agent itself. This integration is essential for facilitating informed decision-making rooted in experiential knowledge. By incorporating probabilistic models to efficiently learn, represent, and reason across various aspects of the agent and its environment, it becomes feasible to equip autonomous robots with cognitive abilities. This empowerment enables them to accurately predict action outcomes based on context, furnishing the agent with essential tools to make well-informed decisions when selecting optimal actions and parameters for their tasks. The presented approach is the first ever to learn and use such comprehensive joint probabilities for a robotic system. Experiments showcase that BayRoB is capable of refining underspecified plans and allow reasoning over arbitrary matters of the agent, the available actions and their parameterizations as well as aspects of the agent’s environment. A browser-based web interface allows the user to investigate the system’s capabilities and reproduce the conducted experiments.