This paper presents a PhD research project aimed at enhancing Human-Robot Interaction (HRI) by empowering non-technical users in both the definition and execution of robot tasks. The project addresses two main objectives: (i) enabling end users, such as domain experts without programming skills, to define, validate, and modify robot tasks through End-User Development (EUD) enhanced by Large Language Models (LLMs); and (ii) improving robot adaptability during collaborative task execution by integrating human cognitive factors (e.g., beliefs, desires, and intentions) into robot decision-making. The findings aim to contribute to the ongoing shift toward more adaptable and human-aware robotic systems.
The field of robotics is currently in a rapid expansion, driven by significant technological advancements
across multiple disciplines. In industrial automation, innovations in control systems, materials, and
technical standards are transforming manufacturing processes. Robots in this domain are capable
of executing highly precise and repeatable operations at high speed over long periods, making them
ideal for tasks such as precision welding, heavy load handling, and complex assembly line operations.
In parallel, the rise of collaborative robots (cobots) is reshaping the way humans and robots interact.
Designed to operate safely alongside human workers without the need for physical barriers, cobots are
enabling more flexible, adaptive, more sustainable and human-centric manufacturing environments [
Furthermore, recent advances in Artificial Intelligence (AI) have introduced a new wave of
computational capabilities to mobile robotics [
Another area undergoing significant transformation is social robotics [
While the robotics field has shown impressive progress and earned substantial attention, challenges persist, particularly in the domains of collaborative and social robotics, where interaction with humans remains a central aspect. Human-Robot Interaction (HRI) occurs at multiple levels: from task definition to real-time task execution and adaptation. These interactions often require a robot to understand and respond to human actions, preferences, and emotional states, for which traditional model-based robotic systems are not well suited.
In collaborative robotics, one key challenge lies in enabling domain experts to define the tasks that robots must perform. Given the increasing versatility of collaborative robots, such systems must be adaptable to frequent changes in production workflows. This demand aligns with the principles of
CEUR Workshop
ISSN1613-0073
End-User Development (EUD) [
Although HRI shares many objectives with Human-Computer Interaction (HCI), some of the main
challenges in building a human-aware collaboration with robots come from the field of autonomous
robotics. These challenges often involve how robots make decisions and interpret their surroundings,
areas traditionally explored through research on autonomous behavior and artificial cognition [
Building on the themes discussed above, this PhD project is structured around two main research objectives. Objective 1 was explored in collaboration with Prof. Daniela Fogli and Prof. Pietro Baroni at the University of Brescia. Objective 2 was developed during a research visit period at CNRS-LAAS in Toulouse (France), under the supervision of Dr. Rachid Alami, and with continued support from Prof. Daniela Fogli.
Starting from the study reported in [
• RQ1: To what extent can domain experts define domain-specific tasks and items through dedicated user interfaces designed for non-programmers? • RQ2: What are the capabilities and limitations of natural language interfaces in supporting users to accurately specify, validate, and refine structured robotic tasks, and how can multimodal interaction enhance this process during task definition?
To address the research questions, several interaction design activities were carried out. Whenever feasible, the investigations focused on use cases in the healthcare domain, particularly on galenic formulations, which are personalized medications manually prepared by pharmacists to meet specific patient needs, such as allergies, tailored dosages, or specific administration forms.
The healthcare domain was selected as a reference context since this PhD project is part of the
Technology for Health doctoral program at the University of Brescia. The interaction design activities
revolved around two main projects:
• Project 1 - Preparation of Personalized Medicines through Collaborative Robots: In [
These projects advance the field of end-user robot programming by introducing a human-centered AI approach that integrates LLMs, multimodal interaction, and digital twins. The resulting workflow empowers domain experts to conceptualize, validate, and refine robot tasks through intuitive, accessible tools, without requiring programming or robotics expertise.
Central to the methodology is the active involvement of end users throughout the design process, ensuring that the environment aligns with their practices, expectations, and cognitive models.
The activities carried out aim to demonstrate that, when properly supported, non-technical users can efectively engage in robot task definition, ultimately enabling more adaptive, usable, and trustworthy robotic systems in real-world environments.
The second objective of the PhD project was to investigate how a robot can efectively collaborate with a human partner by taking into account human beliefs, intentions, and emotional states during task execution. This goal addresses the limitations of rigid model-based approaches, which often fall short when facing the complexity and unpredictability of real-world human behavior.
Key challenges addressed in this objective include: 1. Overcoming the limitations of traditional model-based approaches to robotics when interacting with humans, whose behavior is often non-deterministic and emotionally driven. 2. Maintaining a robust and rigorous task planning framework, even in dynamic and uncertain environments.
To guide this investigation, the following research questions were formulated: • RQ3: How can robots adapt to the unpredictable behavior of human partners in collaborative scenarios, while maintaining task control and ensuring safety during execution?
To address RQ3, a hybrid system architecture was designed that integrates deterministic techniques
(e.g., task planning and rule-based approach) with the flexibility of LLMs. The preliminary
conceptualization was introduced in [
The approach is based on the concept of shared goals between the human and the robot. A task is defined as a cooperative activity aimed at achieving a mutually agreed-upon goal.
The assumed application context is a coaching scenario where the robot assists a person in completing a task composed of several activities. To illustrate this scenario, it can be considered a patient in a care facility who is required to take medications and engage in both cognitive and physical activities. Efective collaboration requires adapting the robot’s actions not only to the requirements of the task but also to the human’s evolving behavior, preferences, and afective state.
The proposed architecture is modular and designed to ensure seamless collaboration, adaptive behavior, and safe execution across the following components: • Natural Language Understanding and Knowledge Acquisition: Interaction begins with natural language input from the user, which is transformed into semantic vector embeddings and stored in a dedicated Vector Database. • Human-Robot Task Synthesizer : This module, based on an LLM, processes user input and contextual data to generate a structured task representation, capturing goals, constraints, and required actions. • Situation Assessment & Human-Robot Task Progress: By integrating prior knowledge, task history, and current context, this module dynamically updates the task plan in response to environmental or human-driven changes. Also, this module is based on an LLM. It ensures continual alignment between system behavior and human expectations. • Event Planner & Next Steps: Responsible for monitoring task progression and deciding subsequent actions, this component checks for discrepancies between expected and actual outcomes, triggering feedback loops when necessary. • Robot Perception & Robot Efector : These modules ensure the system’s responsiveness and safety.
The perception component continuously interprets both the task state and the human’s behavior, leveraging a Visual Language Model (VLM), while the efector module governs physical actions and verbal communication, maintaining alignment with user preferences and situational needs.
The developed architecture contributes to the advancement of HRI by blending deterministic techniques with the contextual flexibility ofered by LLMs. The system is capable of adapting robot behavior in real time, maintaining alignment with human intentions and adjusting to unexpected changes in the environment or in user input.
Empirical case studies confirm the system’s capacity to interpret human-provided goals in natural language, generate coherent and context-aware task plans, dynamically adapt execution strategies in response to ongoing interaction, and ensure safety and reliability despite behavioral uncertainty on the human side.
While the work presented in this paper introduces promising approaches to empowering end users in robot programming and enhancing human-robot collaboration, some limitations remain that ofer implications for future development.
A first limitation relates to the realism of the deployed environments. While the digital twin component introduces a useful simulation layer, it currently simplifies many aspects of the real-world manipulation tasks. Moving forward, a more realistic digital twin, capable of simulating diverse objects and conditions (e.g., diferent objects, workspace constraints, and safety considerations), is needed to better reflect the complexity of actual work settings. Similarly, future user testing should incorporate real objects, such as test tubes and pharmaceutical tools, to more closely replicate the environment in which pharmacists operate and to validate the robustness of the task definition interface under practical constraints.
In the case of the hybrid architecture for collaborative task execution, current experiments rely on partially abstracted or scripted interaction scenarios. A natural next step is to bring this system into more defined and dynamic use cases (e.g., many activities with diferent priorities and schedule constraints, also considering human preferences), where human behavior is less predictable and the system’s adaptability can be more rigorously challenged.
Finally, while recent advances in LLMs ofer new opportunities for natural and flexible interaction, they must be approached with critical awareness. In the projects presented in this paper, LLMs are never used as-is; instead, they are systematically complemented by additional technologies and methodologies designed to interpret, verify, and ground user input. This hybrid strategy is essential to mitigate the known limitations of LLMs, such as inconsistencies or safety concerns, and to ensure that human-robot collaboration remains robust, reliable, and context-aware.
Overall, this work puts efort into enabling more human-aware and intelligent robotic systems. By demonstrating the feasibility of key concepts, this work provides a solid foundation for future advancements in the field. Building on these results, the next steps will involve refining the approach through real-world testing, enhancing simulation environments, and fostering continuous user engagement to ensure robust and practical deployment.
Declaration on Generative AI During the preparation of this work, the author used ChatGPT and Grammarly in order to: grammar and spelling check, paraphrase, and reword. After using these services, the author reviewed and edited the content as needed, thus, he takes full responsibility for the publication’s content.
Funding The PhD scholarship is co-funded by the Italian Ministry, Piano Nazionale di Ripresa e Resilienza (PNRR) and Antares Vision.