Till Örebro universitet

Autonomous systems are increasingly being deployed across various real-world domains, such as fleets of self-driving vehicles, robotic warehouses, and delivery services using unmanned aerial vehicles. These systems are required to operate with high reliability and predictability, to adapt continuously to changing conditions, and to remain accountable to human supervisors. To achieve these objectives, autonomous systems need explicit, formal representations of their behavior that facilitate task planning, system verification, and human oversight.

In current industrial practice, such representations, whether for task-level control or action planning, are typically engineered manually. While handcrafted representations can be precise, their development is labor-intensive and difficult to scale. Learning-based approaches offer a promising alternative by extracting behavioral representations from execution data. However, they often make unrealistic assumptions, such as access to simulated environments or large volumes of high-quality training data. Moreover, they fail to simultaneously achieve all the three critical objectives, that is reliability, adaptability, and interpretability. Therefore, there is a clear need for methods capable of efficiently learning accurate, interpretable representations under realistic conditions.

In this thesis, we address the problem of learning interpretable representations of system behavior from execution traces - sequences of observed actions and state transitions generated during the operation of autonomous systems. Learning from such traces is appealing because they are readily available from system logs and provide direct evidence of how a system behaves in realistic, often complex environments. The overarching goal is to derive representations that not only support automated planning but also enhance human understanding and oversight.

Two distinct types of behavior representation are explored: Behavior Trees (BTs) and STRIPS-style planning domains. For each, a novel method to automatically construct representations from execution traces is proposed. Specifically, for BTs, we introduce a method that combines Boolean logic, leveraging algorithms originally developed in circuit theory, with decision tree learning to induce structured, interpretable behavior representations. To assess the interpretability of BTs, a user study is conducted to examine how such representation sare perceived by human users. The study identifies key features that influence user comprehension, contributing empirical evidence to a space that has traditionally lacked systematic analysis. Furthermore, a structured evaluation method for BTs along with quality metrics and design principles is presented, addressing the current lack of guidance for assessing BT quality beyond functional performance.

For STRIPS-style domains, we introduce a novel learning framework to construct symbolic action representations directly from execution traces, even in the presence of noise. In addition to the learning algorithm, a systematic methodology is proposed for evaluating learned planning domains through structural and task-based analysis, thereby addressing a critical gap in current practice and thus responding to the growing need for rigorous assessment methods.

The results demonstrate that it is possible to extract interpretable representations of autonomous behavior from noisy data. The proposed methods enable the transition from raw execution traces to structured representations that can support planning, validation, and human-in-the-loop systems. By advancing methods for learning, interpreting, and evaluating learned behavior representations, this work contributes to the development of autonomous systems that are both operationally effective and intelligible to human stakeholders.

Automated planning (AI planning) solves complex sequential decision-making problems by searching for action sequences given a domain definition that models the actions a system can take. However, hand-crafting action models for complex problems is often challenging and cumbersome even for experts. To address this, we present Lifted Action Model Learning (LAML), a novel approach for learning planning domains from plan traces obtained by noisy observations of environment states. LAML integrates data lifting, decision-tree learning, and logic induction to derive abstract action models and generate planning domain representations. We evaluate our approach on 23 benchmark domains from the International Planning Competition, comparing its performance to state-of-the-art methods. Our evaluation considers multiple criteria, including domain reconstruction through comparison with reference domains, plan generation feasibility, comparison with historical plans, and plan validation success rate. Experimental results demonstrate that LAML not only reconstructs more accurate action models but also exhibits strong robustness to noise.

Behavior trees (BTs) are increasingly popular in the robotics community. Yet in the growing body of published work on this topic, there is a lack of consensus on what to measure and how to quantify BTs when reporting results. This is not only due to the lack of standardized measures, but due to the sometimes ambiguous use of definitions to describe BT properties. This work provides a comprehensive overview of BT properties the community is interested in, how they relate to each other, the metrics currently used to measure BTs, and whether the metrics appropriately quantify those properties of interest. Finally, we provide the practitioner with a set of metrics to measure, as well as insights into the properties that can be derived from those metrics. By providing this holistic view of properties and their corresponding evaluation metrics, we hope to improve clarity when using BTs in robotics. This more systematic approach will make reported results more consistent and comparable when evaluating BTs.

One of the difficulties of using AI planners in industrial applications pertains to the complexity of writing planning domain models. These models are typically constructed by domain planning experts and can become increasingly difficult to codify for large applications. In this paper, we describe our ongoing research on a novel approach to automatically learn planning domains from previously executed traces using Behavior Trees as an intermediate human-readable structure. By involving human planning experts in the learning phase, our approach can benefit from their validation. This paper outlines the initial steps we have taken in this research, and presents the challenges we face in the future.

The increased popularity of Behavior Trees (BTs) in different fields of robotics requires efficient methods for learning BTs from data instead of tediously handcrafting them. Recent research in learning from demonstration reported encouraging results that this letter extends, improves and generalizes to arbitrary planning domains. We propose BT-Factor as a new method for learning expert knowledge by representing it in a BT. Execution traces of previously manually designed plans are used to generate a BT employing a combination of decision tree learning and logic factorization techniques originating from circuit design. We test BT-Factor in an industrially-relevant simulation environment from a mining scenario and compare it against a state-of-the-art BT learning method. The results show that our method generates compact BTs easy to interpret, and capable to capture accurately the relations that are implicit in the training data.

This short paper presents an ongoing industrial PhD project focused on learning domain models in the transportation industry. We seek to develop a two-phase learning process that leverages Behavior Trees to involve human planning experts in the learning loop. In this paper, we outline our research questions, discuss the initial steps we have taken in our research, and highlight the challenges we expect to face moving forward.