Autonomous systems often operate under complex task constraints that involve both spatial and temporal aspects, such as "the robot must inspect Area 1 for 3 minutes, then Area 2 for 5 minutes, and finally reach the goal while maintaining a speed below 5 m/s at all times". Signal Temporal Logic (STL) is a formal, logic-based framework to specify these objectives. While some of the existing research has focused on using STL to generate robot trajectories that satisfy predefined constraints, this paper tackles the inverse problem: given successful trajectories, how can we infer the underlying STL constraints they adhere to? We address this problem by using Parametric STL (PSTL), where domain experts define template STL formulas with certain values left as symbolic parameters (e.g., "the robot must inspect Area 1 for t_1 minutes"). The goal is to determine the parameter values that best replicate the expert trajectories. To this end, we introduce a framework that uses Bayesian Optimization (BO) to identify the unknown parameters of the PSTL formulas. This approach leverages BO’s ability to explore and exploit parameter spaces with minimal evaluations, ensuring that the learned PSTL parameters closely match expert demonstrations while preserving the interpretability and structure of the original task specifications. Our experimental results demonstrate that the proposed framework can accurately learn PSTL parameters with an error of less than 10%, although certain edge cases still present challenges. Additionally, we explore the amount of expert data required to effectively learn these parameters and show that even a small set of 20-25 expert trajectories can produce accurate results that capture key task behaviors.

Geotechnical engineering faces challenges in subsurface characterization due to reliance on sparse site investigations and empirical correlations, which limit the accuracy of soil property estimations. This study addresses this issue by leveraging machine learning (ML) methods to predict cone penetration test (CPT) measurements using data from driven cast in-situ (DCIS) pile installations. A comprehensive dataset, including cone resistance, sleeve friction, and friction ratio measurements, was used to train and evaluate ML models—Random Forest (RF), eXtreme Gradient Boosting (XGB), and TabNet. Among these, XGB trained at a 30 cm depth resolution demonstrated a strong balance of predictive accuracy, efficiency, and granularity, achieving R2 values exceeding 80% for cone resistance predictions even with limited data availability. By improving predictions at unseen locations, the study showcases the feasibility of integrating ML models into geotechnical workflows to enhance real-time decision-making, optimize construction practices, and support the development of digital twins for pile-supported structures.
Interpretability methods, specifically SHAP values, provided insights into feature significance and highlighted the critical roles of depth, spatial coordinates, machine-related features and representative sampling strategies in achieving reliable predictions. These findings demonstrate ML’s potential to align predictive capabilities with engineering principles, bolstering confidence in its adoption for practical applications.

Despite rapid advancements in Large Language Models (LLMs), they often produce hallucinated or detrimental outputs, necessitating alignment with human preferences. We address these challenges by introducing Step Chain-of-Thought (SCoT) to enhance semantic understanding by breaking down complex instructions. Additionally, we combine Direct Preference Optimization (DPO) with Low-Rank Adaptation (LoRA) to improve alignment with user intent. DPO optimizes outputs based on human feedback, while LoRA, alongside careful tuning of learning rates and beta values, mitigates repetition issues seen with DPO alone. Our findings show that models fine-tuned with DPO with LoRA achieve superior alignment compared to those using only Supervised Fine-Tuning (SFT). However, automated evaluators like LLM-as-a-Judge struggle with nuanced SCoT assessments, underscoring the necessity of human evaluation for capturing the complexities of alignment. In task alignment for robotics, Full Fine-Tuning (FFT) excels in familiar tasks, while LoRA significantly improves adaptability to new scenarios, increasing the robustness. Moreover, combining ground truth with synthetic data, especially when using LoRA, achieves a balance between accuracy and adaptability, revealing the limitations of relying solely on synthetic data. These conclusions highlight the critical importance of well-aligned datasets, fine-tuning strategies, and careful parameter tuning for LLM alignment.

To ensure safe operation of autonomous vehicles (AVs), trajectory planners should account for occlusions. These are areas invisible to the AV that might contain vehicles. Set-based methods can guarantee safety by calculating the reachable set, which is the set of possible states, for each potentially hidden vehicle. A recently published method proved in simulation experiments to reduce the cautiousness by reasoning about these occluded areas over time, assuming perfect input data. We present a novel algorithm that uses this reasoning and is applicable on a real AV with its accompanying uncertainties and imperfect sensor data. The uncertainties include sensor errors and noise, computation and communication delays and control errors in the trajectory following. This is achieved by modelling the error distributions and accounting for them in the calculations, where the confidence interval for each error is exposed as a setting. Experiments indicate that our algorithm can reduce the traversal time through an intersection by 2.2 seconds with reasoning. An ablation study of the different error measures shows that the errors in the construction of the field of view (FOV) limit the performance the most. Reducing the errors in the FOV construction is therefore the most important recommendation, besides making the method interaction-aware.

Autonomous mobile robots are increasingly performing tasks in our daily environments, e.g., cleaning offices or order picking in supermarkets.
In such human-populated scenarios, it is crucial that these robots always navigate safely when performing their task.
Existing state-of-the-art methods can ensure safety, but often require deterministic motions of humans and may lead to conservative behavior in dense human-populated environments. This study investigates the use of time-varying control barrier functions (CBFs) and time-based rapidly exploring random trees (RRTs) to combine local safety with global task performance.
The proposed new cost function improves the trade-off of safety and task progress in densely populated scenarios.
Our results show that time-varying CBFs can perform better in terms of both task performance and safety compared to normal CBFs. Furthermore, our real-world robot experiments validate our approach to a physical nonholonomic robot.

Autonomous robots are increasingly improving at performing navigation tasks, however they will likely fail at some point or not perform as intended due to uncertainties or unforeseen situations in the real world. In such scenarios, explaining the robot's behaviour to humans is crucial to build trust and resolve potential issues. Recently, large language models (LLMs) have shown great potential in analysing robot log data, e.g., obtained in Robot Operating System (ROS), and providing users with useful explanations. Yet, these models often fail to consistently generate high quality answers. This study develops an approach using parameter-efficient fine-tuning (PEFT) to improve explanations generated by LLMs and tailoring them towards a target audience (expert, non-expert) and preferred lengths (short, medium, long). We collected ROS log data from the TIAGo robot in simulation, combined them with user questions, and corresponding answers generated using GPT-4o to create a dataset for fine-tuning Mistral 7B with PEFT. Furthermore, we use a panel of LLMs (GPT-4o, Mistral-Large, Llama3-8B) as judges to evaluate these explanations based on quality criteria and user study (N=17) to validate these results on a group of roboticists. Our findings show that personalisation significantly improves both the suitability of explanations, with personalised answers consistently outperforming non-personalised ones. Furthermore, tailored explanations achieved higher clarity and user understanding. Additionally, a single feedback loop iteration using textual feedback from LLMs further enhanced explanation relevance and contextual quality, demonstrating the value of iterative improvement in explainability systems, despite minor trade-offs in other criteria.

Behaviour trees (BTs) serve as a powerful hierarchical structure for task execution, simplifying complex tasks but posing challenges in their manual design. The automatic generation of BTs addresses this concern, yet often lacks robust failure recovery options. This study presents Failure Recovery with Ontologically Generated Behaviour Trees (FROGBT), a novel approach bridging this gap by integrating ontological reasoning into the process of automatically generating BTs. This integration establishes a profound link between an agent’s knowledge and its capabilities, offering contextual insights into the agent’s skills. FROGBT enhances skill representation for planning and recovery. The approach’s effectiveness is indicated by its efficiency compared to the state-of-the-art framework for skill-based control, SkiROS, in a similar task. It showcases generality, uniting diverse skills, developed by various engineers, for recurring tasks, and introduces innovative failure recovery strategies. FROGBT highlights ontological reasoning’s potential to enhance BT generation with context-awareness and reasoning abilities, paving the way for future research on failure recovery concepts in generated BTs.