Autonomous systems often operate in multiple operational modes, each with its own set of safety constraints. Safely transitioning between modes, for example, from a high-speed survey mode to a close-inspection mode, presents a significant challenge. This paper introduces the ‘Elastic’ Model Predictive Safety Filter (EPSF), a novel control framework that enables provably safe, dynamic transitions between different operational modes. The EPSF extends the Predictive Safety Filter by incorporating a soft-constrained multimodal Model Predictive Control formulation. In transition between modes, the controller’s cost function penalises deviations from the constraints of a desired target mode whilst strictly enforcing the constraints of the currently operational mode. This objective steers the system state towards a region where a transition is feasible. The convergence of slack variables to zero serves as an online, verifiable certificate that this transition region has been reached, guaranteeing recursive feasibility and asymptotic convergence. A convex implementation is provided for linear systems, and simplifications of the sufficient conditions for convergence are discussed. The effectiveness of the EPSF is demonstrated through numerical simulation of an unstable linear system and a nonlinear kinematic car model.

Unmanned aerial vehicles (UAVs) play an increasingly vital role in wilderness search and rescue (SAR) operations by enhancing situational awareness and extending the capabilities of human teams. Yet, a lack of standardized benchmarks has impeded the systematic evaluation of single- and multi-agent path-planning algorithms. This paper introduces an open-source dataset and evaluation framework to address this gap. The framework comprises 60 geospatial scenarios across four distinct European environments, featuring high-resolution probability maps. We present a lost person probabilistic model derived from statistical models of lost person behavior. We provide a suite of tools for evaluating search paths against four baseline methods: Concentric Circles, Pizza Zigzag, Greedy, and Random Exploration, using three quantitative metrics: Accumulated probability of detection, time-discounted probability of detection, and lost person discovery score. We provide an evaluation framework to facilitate the comparative analysis of single- and multi-agent path-planning algorithms, supporting both the baseline methods presented and custom user-defined path generators. By providing a structured and extensible framework, this work establishes a foundation for the rigorous and reproducible assessment of UAV search strategies in complex wilderness environments.

As the size and mass of orbital structures increase, recent research into the testing and development of technology for assembling these large structures has led to the development of novel techniques. One such method involves walking robots and strategically placed logistical depots. This paper describes the development of a high-level discrete model with two distribution algorithms for the logistical depots and two competing assembly sequence planning strategies. The results show that the alteration of different variables and assembly strategies significantly influences assembly efficiency, with some assembly strategies specifically suiting certain metrics. It also appears that depot distribution has a more complex relationship with the optimisation of the assembly process.

Persistent monitoring using robot teams is of interest in fields such as security, environmental monitoring, and disaster recovery. Performing such monitoring in a fully on-line decentralised fashion has significant potential advantages for robustness, adaptability, and scalability of monitoring solutions, including, in principle, the capacity to effectively adapt in real-time to a changing environment. We examine this through the lens of multi-robot patrol, in which teams of patrol robots must persistently minimise time between visits to points of interest, within environments where traversability of routes is highly dynamic. These dynamics must be observed by patrol agents and accounted for in a fully decentralised on-line manner. In this work, we present a new method of monitoring and adjusting for environment dynamics in a decentralised multi-robot patrol team. We demonstrate that our method significantly outperforms realistic baselines in highly dynamic scenarios, and also investigate dynamic scenarios in which explicitly accounting for environment dynamics may be unnecessary or impractical.

Robot swarms have shown great potential for exploration of unknown environments, utilizing simple robots with local interaction and limited sensing. Despite this, complex indoor environments can create issues for reactive swarm behaviours where specific paths need to be travelled and bottlenecks are present. In this paper we present our social exploration algorithm which allows the swarm to decide between different options of swarm behaviours to search randomly generated environments. Using a “happiness” measure, agents can reason over the performance of different swarm behaviours, aiming to promote free movement. Agents collaborate to share opinions of different behaviours, forming teams which are capable of adapting their exploration to any given environment. We demonstrate the ability of the swarm to explore complex environments with minimal information and highlight increased performance in relation to other swarm behaviours over 250 randomly generated environments.

We propose a method to capture the handling abilities of fast jet pilots in a software model via reinforcement learning (RL) from human preference feedback. We use pairwise preferences over simulated flight trajectories to learn an interpretable rule-based model called a reward tree, which enables the automated scoring of trajectories alongside an explanatory rationale. We train an RL agent to execute high-quality handling behaviour by using the reward tree as the objective, and thereby generate data for iterative preference collection and further refinement of both tree and agent. Experiments with synthetic preferences show reward trees to be competitive with uninterpretable neural network reward models on quantitative and qualitative evaluations.

Unmanned aerial vehicles (UAVs) have become crucial for various applications, necessitating reliable and time-constrained performance. Multi-UAV solutions offer advantages but require effective coordination. Traditional coverage path planning methods overlook uncertainties and individual UAV failures. To address this, reliability-aware multi-UAV coverage path planning methods optimise task allocation to maximise mission completion probabilities given a failure model. This paper presents an experimental validation of the reliability-aware approach, specifically an approach using a Greedy Genetic Algorithm (GGA). We evaluate the GGA performance in real-world environments, comparing mission reliability to computed reliability and comparing it against a traditional multi-UAV methods. The experimental validation demonstrates the practical viability and effectiveness of the reliability-aware approach, showing significant improvement in mission reliability despite the inevitable mismatch between real and assumed failure models.

This paper describes the partial automation of drones (also referred to as uncrewed aerial vehicles, UAVs, or aerial robots) in populated areas within the visual line-of-sight of their pilot. Mission responsiveness is improved by reducing the number of human crew members and avoiding the need for area clearance, while carefully managing the workload of those remaining to ensure no compromise on safety. The work employs a system-centric approach with regard to integrating human and automation tasks based on their capabilities and use of standard procedures, whilst prioritizing the predictability and simplicity of the overall system. Safety claims about the proposed system are posed and rigorously analyzed through a structured safety case. The proposed system is applied to a bridge inspection case study with simulation results and scenario analysis.

Model-based reinforcement learning (RL) algorithms do not typically consider environments with multiple dynamic modes, where it is beneficial to avoid inoperable or undesirable modes. We present a model-based RL algorithm that constrains training to a single dynamic mode with high probability. This is a difficult problem because the mode constraint is a hidden variable associated with the environment's dynamics. As such, it is 1) unknown a priori and 2) we do not observe its output from the environment, so cannot learn it with supervised learning. We present a nonparametric dynamic model which learns the mode constraint alongside the dynamic modes. Importantly, it learns latent structure that our planning scheme leverages to 1) enforce the mode constraint with high probability, and 2) escape local optima induced by the mode constraint. We validate our method by showing that it can solve a simulated quadcopter navigation task whilst providing a level of constraint satisfaction both during and after training.

Recent efforts to learn reward functions from human feedback have tended to use deep neural networks, whose lack of transparency hampers our ability to explain agent behaviour or verify alignment. We explore the merits of learning intrinsically interpretable tree models instead. We develop a recently proposed method for learning reward trees from preference labels, and show it to be broadly competitive with neural networks on challenging high-dimensional tasks, with good robustness to limited or corrupted data. Having found that reward tree learning can be done effectively in complex settings, we then consider why it should be used, demonstrating that the interpretable reward structure gives significant scope for traceability, verification and explanation.

Autonomous systems increasingly are integrated into larger, connected, and hybrid (Human-Machine) systems of systems, making them complex systems - which are hard to design and predicting emergent behaviour is difficult. These issues are faced increasingly across civil and military applications, both in the UK and NATO. A holistic approach is needed to fully quantify them. Working as a partnership between industry and academia has provided greater freedom to apply innovative technologies in the context of relevant use cases. This paper presents some tools and methods we have used in our research and development to support this approach and address the challenges of deploying autonomous systems in the future. We discuss the use of simulations and how they can support every step of the process, from academic experiments to digital twins; where the right level of fidelity is needed at different times to give maximum benefit. The use of a common simulation platform to align control design exploration with human factors research is discussed, enabling questions of human-machine teaming and trust. We highlight how foundational research on: architecture and modelling, network topology, decision making processes and human interactions impact on the overall development of a system. Included are our lessons identified from this partnership.

We introduce a data-driven, model-agnostic technique for generating a human-interpretable summary of the salient points of contrast within an evolving dynamical system, such as the learning process of a control agent. It involves the aggregation of transition data along both spatial and temporal dimensions according to an informationtheoretic divergence measure. A practical algorithm is outlined for continuous state spaces, and deployed to summarise the learning histories of deep reinforcement learning agents with the aid of graphical and textual communication methods. We expect our method to be complementary to existing techniques in the realm of agent interpretability.

One of the major challenges of integrating AI in UAV control is the ability of the pilot to effectively interact with the system.Inspired by process-control and vigilance devices, this work employs a predictable and deterministic form of AI to enable single-crew (i.e.pilot) operation in an unconstrained dynamic environment.A multi-user simulation environment was developed in Unity to validate the mission and to train and evaluate pilot-automation interactions.Preliminary simulation results for an inspection case study at the Clifton Suspension Bridge are reported.

Through the use of autonomy Unmanned Aerial Vehicles (UAVs) can be used to solve a range of of multi-agent problems that exist in the real world, for example search and rescue or surveillance. Within these scenarios the global objective might often be better achieved if aspects of the problem can be optimally shared amongst its agents. However, in uncertain, dynamic and often partially observable environments centralised global-optimisation techniques are not achievable. Instead, agents may have to act on their own belief of the world, making the best decisions independently and potentially myopically. With multiple agents acting in a decentralised manner how can we discourage competitive behaviour and instead facilitate cooperation. This paper focuses on the specific problem of multiple UAVs simultaneously searching for tasks in an environment whilst efficiently routing between them and ultimately visiting them. This paper is motivated by this idea that collaboration can be simple and achieved without the need for a dialogue but instead through the design of the individual agent’s behaviour. By focusing on what is communicated we expand the use of a single agent behaviour. Which through minor modifications can produce distinct agents demonstrating independent, collaborative and competitive behaviour. In particular by investigating the role of sensor and communication ranges this paper will show that increased sensor ranges can be detrimental to system performance, and instead the simple modelling of nearby agents’ intent is a far better approach.

Whilst swarms have potential in a range of applications, in practical real-world situations, we need easy ways to supervise and change the behaviour of swarms to promote robust performance. In this paper, we design artificial supervision of swarms to enable an agent to interact with a swarm of robots and command it to efficiently search complex partially known environments. This is implemented through artificial evolution of human readable behaviour trees which represent supervisory strategies. In search and rescue (SAR) problems, considering uncertainty is crucial to achieve reliable performance. Therefore, we task supervisors to explore two complex environments subject to varying blockages which greatly hinder accessibility. We demonstrate the improved performance achieved with the evolved supervisors and produce robust search solutions which adapt to the uncertain conditions.

This work considers dynamically feasible point-to-point trajectory generation problem for a quadrotor flying through a constrained planner region referred as window (narrow gap). A four parameter logistic (4PL) curve is investigated as a prospective candidate and closed-form conditions are derived on the 4PL design parameters to satisfy the window traversability and vehicle dynamic feasibility constraints. A hierarchical approach first computes a dynamically feasible design parameter set for decoupled trajectory components and then obtains a solution set satisfying 3-D axis-coupled window traversability conditions. Numerical examples with a comparative study are presented to validate the analytical findings that highlight the quick computation of the dynamically feasible window traversing trajectories in complex window scenarios.

Wheeled robots typically depend on a mix of Wheel Odometry (WO) and Visual Odometry (VO) for localization. This includes inference of the slip of the wheels, representing different soil interactions. This paper investigates how slip predictions derived from forward vision or drive current, for example, can be fused with WO and VO. Our solution is based on the Extended Kalman Filter (EKF), using either point slip measurements, a slip prediction model, or a slip profile as a state. The investigation also compares different VO measurement periods, studying how VO effort trades with accuracy for each slip prediction scheme. The solutions all provide improved localization accuracy and hint at the intriguing possibility of slip-based SLAM.

A new strategy is proposed for scheduling Visual Odometry (VO) measurements for wheeled ground vehicles. Rather than having a fixed interval or distance between image acquisitions, we propose to trigger VO based on covariances from an Adaptive Extended Kalman Filter. The adopted model uses process noise to drive wheel slip estimation, which, when correctly identified, can be used with Wheel Odometry to provide frequent position estimates. When more dynamic terrain is detected, more VO measurements are scheduled to maintain localization accuracy. On the other hand, when the terrain is stable, VO usage is limited. The system is validated in a simple one-dimensional case using data captured during field trials using a representative rover. The results are promising as trajectories that were subjected to large errors are corrected.

This work considers the problem of generating computationally efficient quadrotor motion primitives between a given pose (position, velocity, and acceleration) and a goal plane in the presence of obstacles. A new motion primitive tool based on the logistic curve is proposed and a closed-form analytic approach is developed to satisfy constraints on starting pose, goal plane, velocity, acceleration, and jerk. The geometric obstacle avoidance problem is represented as a combinatorial set problem and a heuristic approach is proposed to accelerate the solution search. Numerical examples are presented to highlight the fast motion primitive generation in multi-obstacle pose-to-plane scenarios.