Language-conditioned local navigation requires a robot to infer a nearby traversable target location from its current observation and an open-vocabulary, relational instruction. Existing vision-language spatial grounding methods usually rely on vision–language models (VLMs) to reason in image space, producing 2D predictions tied to visible pixels. As a result, they struggle to infer target locations in occluded regions, typically caused by furniture or moving humans. To address this issue, we propose BEACON, which predicts an ego-centric Bird’s-Eye View (BEV) affordance heatmap over a bounded local region including occluded areas. Given an instruction and surround-view RGB-D observations from four directions around the robot, BEACON predicts the BEV heatmap by injecting spatial cues into a VLM and fusing the VLM’s output with depth-derived BEV features. Using an occlusion-aware dataset built in the Habitat simulator, we conduct detailed experimental analysis to validate both our BEV space formulation and the design choices of each module. Our method improves the accuracy averaged across geodesic thresholds by 22.74 percentage points over the state-of-the-art image-space baseline on the validation subset with occluded target locations.

Autonomous mobile robots have become increasingly capable over the past decades, enabling their use in domains such as logistics, agriculture, and healthcare. This thesis contributes to a collaborative project between the Dutch National Police and the Delft University of Technology, focusing on search and rescue operations in the security domain.

In these safety-critical missions, robots are expected to support human first responders by exploring unknown and potentially hazardous environments, including buildings and natural terrains. To reduce the cognitive and operational burden on human operators, robots must navigate cluttered spaces both autonomously and safely. In the context of this thesis, autonomy refers to onboard decision-making and safety to avoiding collisions despite uncertainties like sensor noise or environmental effects.

This thesis explores autonomous and safe navigation using model predictive control (MPC), a planning and control strategy that predicts future behavior based on a system model and optimizes actions accordingly. MPC’s strength lies in its ability to handle constraints such as actuator limits and obstacle avoidance. However, applying MPC in real-world navigation presents several challenges, which this thesis addresses through the following contributions.

Navigating a mobile robot through cluttered environments requires it to follow dynamic trajectories while avoiding obstacles. Although MPC is well-suited for this task, recent methods for collision-free trajectory tracking often rely on complex mathematical formulations that can be difficult to interpret and apply. Chapter 2 aims to bridge this gap by offering a structured, step-by-step guide. It explains how to model the robot’s navigation problem, formulate the corresponding MPC problem, and establish performance and safety guarantees.

The chapter presents three formulations: a nominal MPC approach for ideal conditions, a robust MPC formulation that accounts for bounded disturbances, and a robust output-feedback MPC formulation that additionally handles measurement noise. Each formulation is supported by theoretical insights and practical considerations. While not exhaustive, the guide is intended to support researchers and practitioners in implementing MPC-based navigation under varying levels of uncertainty.

To enable autonomy, the MPC formulations introduced in Chapter 2 must operate in real time. This is challenging because MPC relies on solving an optimization problem at each time step, which can be computationally demanding. Planning long-term trajectories and computing control commands at high frequency on embedded hardware is especially difficult.

Chapter 3 addresses this by introducing a hierarchical MPC (HMPC) framework that separates planning and control into two layers. The planning MPC handles long-term trajectory generation at a lower frequency, while the tracking MPC focuses on short-term execution at a higher frequency. This separation allows the use of complex nonlinear models in both layers without compromising real-time performance. The tracking MPC layer, freed from long-term planning, can focus on precise tracking, improving stability and responsiveness.

The HMPC framework also includes a method for generating consistent collision avoidance constraints. Its effectiveness is demonstrated through simulations and lab experiments, showing safer and approximately four times faster goal-reaching compared to a single-layer MPC approach.

While the HMPC framework improves autonomy, it leverages a nominal MPC formulation that assumes perfect models and accurate state data. In practice, mobile robots operate in the presence of model uncertainties and noisy measurements, which can lead to constraint violations such as collisions.

Chapter 4 extends the HMPC framework to address these issues by incorporating robust output-feedback MPC into the tracking layer. This extension, called robust output-feedback hierarchical MPC (ROHMPC), provides formal safety guarantees even in the presence of disturbances and measurement noise. Synthesizing the ROHMPC scheme requires knowledge of uncertainty bounds, which are typically unknown.

To overcome this, the chapter introduces an efficient and modular pipeline that estimates these bounds from experimental data, performs necessary offline computations, calibrates constraint tightening to reduce conservatism, and implements the complete control scheme. The pipeline is released as an open-source software package to support reproducibility and future research. Using this pipeline, the chapter demonstrates the successful validation of the ROHMPC framework properties on a simulated quadcopter platform in Gazebo, with reproducible results.

The HMPC and ROHMPC frameworks address autonomy and safety respectively, but successful deployment in real-world scenarios also depends on the reliability of the system. A key aspect of reliability is reproducibility: the ability to consistently generate similar results.

Chapter 5 explores this concept in the context of robotics, defining reproducibility and analyzing how it applies to the hardware-software setups used in earlier chapters. The HMPC framework satisfies method reproducibility, meaning its implementation can be consistently reproduced. However, due to asynchronous processes and non-deterministic code components, it does not fully achieve results reproducibility.

In contrast, the ROHMPC framework satisfies both method and results reproducibility, reinforcing the credibility of the framework. By raising awareness of reproducibility challenges and offering practical insights, this chapter aims to support the development of more robust and trustworthy robotic systems.

All results presented in this thesis have been made publicly accessible through submissions to peer-reviewed venues, an open-access preprint server, and the release of open-source software packages. These results highlight the effectiveness of hierarchical MPC in both simulation and laboratory settings, and demonstrate how formal safety guarantees can be achieved under uncertainty.

To support future research, Chapter 6 summarizes the key contributions and outlines several promising directions for further exploration. These include extending the proposed algorithms to 3D environments, integrating onboard sensing for autonomous outdoor navigation, and incorporating data-driven methods to reduce conservatism.

As robots leave factory floors and deploy into unstructured human environments, they must navigate safely and efficiently alongside people and other autonomous systems. In these dynamic settings, decisions are inherently interdependent: a robot's optimal action depends heavily on how others will respond, and vice versa. Anticipating and actively shaping these responses is central to competent multi-agent behavior. However, this interaction unfolds under significant uncertainty, with limited prior knowledge of other agents' objectives, constrained sensing capabilities, and tight computational limits.

This dissertation addresses these fundamental challenges by framing robot motion planning for interactive scenarios through the lens of non-cooperative game theory. This perspective provides a principled mathematical framework for modeling multiple self-interested decision-makers who act simultaneously with partially aligned objectives. Focusing on environments where a single controlled robot interacts with uncontrolled agents whose intents are not known a priori, this dissertation develops a comprehensive suite of game-theoretic tools. These tools enable robots to infer underlying intents from observations and generate motion plans that capture the complex interdependence of self-interested decision-making.

The core contributions of this dissertation span intent inference, real-time adaptation, uncertainty-aware planning, computational efficiency, and complex non-smooth dynamics.

First, we formalize the problem of learning unknown intents from observed past behavior as an inverse game. By casting this as maximum-likelihood estimation with equilibrium constraints, our transcription jointly estimates game parameters, hidden states, and future decisions, significantly improving inference accuracy. Second, we tightly integrate these inverse games with online planning. We propose a novel solution technique handling inequality constraints with a first-order update rule for amortized inference, yielding a game-theoretic planner that dynamically adapts to evolving intent estimates.

Third, we address situations demanding explicit reasoning over a distribution of possible intents. We introduce contingency games, an uncertainty-aware planning technique that jointly generates multi-hypothesis predictions of others alongside conditional plans for the robot. By explicitly anticipating future information gains, this approach seamlessly bridges the gap between conservatively ignoring uncertainty and assuming it will never resolve. Fourth, to alleviate the substantial computational burden of online game-theoretic planning, we introduce an amortized solver for mixed strategies. An offline model learns to propose dynamically feasible trajectory candidates, while a discrete game solved online rapidly computes competitive mixed Nash equilibria.

Finally, we tackle interaction domains with inherently non-smooth dynamics, such as multi-agent manipulation, where constraints are not continuously differentiable. We propose a data-driven approach leveraging probabilistic inference and generative diffusion models. This blends learning from single-agent demonstrations with reasoning about joint multi-agent costs, discovering collaborative strategies without requiring massive multi-agent datasets.

In summary, this dissertation advances interactive motion planning by equipping robots to accurately infer intents, act safely under uncertainty, and navigate complex interactions. These algorithmic contributions are extensively validated via simulation and ground robots across autonomous driving, mobile navigation, and multi-agent manipulation, accompanied by open-source libraries to accelerate future research.

Driving in urban environments is a challenging task, even for us humans. There is a variety of different traffic participants including other drivers, cyclists and pedestrians who each have their own unique behaviors. In order to navigate around them safely it is not enough to have detected their presence. We also need to reason about what they might do. However, we cannot always be certain of others’ course of action as we do not know which route they are taking, how risk-averse they are or how aware they are of their surroundings, along with a number of other psychological factors. Hence, we also need to reason how likely it is that they will take a particular course of action. For autonomous vehicles to navigate in urban environments with other traffic participants, we need to imbue them with the capacity to reason about what other participants may do. The main goal of this thesis is to provide contributions in the development of probabilistic trajectory prediction models which accurately capture the distribution over possible future trajectories of other traffic participants.....

Collaborative transportation and manipulation of cable-suspended loads by multiple UAVs offer a promising way for expanding UAVs’ role in heavy-lifting operations. Existing approaches for collaborative aerial manipulation of a payload along a reference trajectory typically rely either on centralized control architectures or on reliable inter-agent communication. In this work, we propose a novel machine learning–based method for decentralized kinodynamic planning that operates effectively under partial observability and without inter-agent communication. Our method leverages imitation learning to train a decentralized homogenous student policy for each UAV by imitating a centralized kinodynamic motion planner which has access to privileged global observations. The student policy uses physics-informed neural networks that respect the derivative relationships of motion to generate trajectory that are step-wise consistent and guaranteed to be kinematically feasible. During training, the student policies utilize the full trajectory generated by the teacher policy, leading to improved sample efficiency. Therefore, the student policy can be trained in under two hours on modest hardware. We validate our method in both simulation and real-world environments to follow an agile reference trajectory, demonstrating performance comparable to that of centralized approaches.

We propose a framework that enables an agent to plan effectively under interaction uncertainty in decentralized multi-agent task and motion planning settings for cooperative manipulation tasks. In decentralized systems, each agent computes plans locally, based on local observations and assumptions on the other agents. Decentralization leads to each agent being uncertain about the actions and behavior of the other agents in the scene. We refer to this as ”interaction uncertainty”, as it arises from the implicit interaction between agents. This stems from limitations of the perception system and the inherent ambiguity of actions, meaning that an agent is not certain about the action being performed by the other agent(s). In other words, it is not certain if the other agents’ observed actions will obtain desired outcomes. In addition, an agent cannot predict the future behavior of the other agents or how this behavior is influenced by its own actions. The proposed framework addresses these two levels of interaction uncertainty in two-agent settings by leveraging the partial observability of the other agent’s short-term intent through ego’s perception system, and modeling the behavior of the other agent through an interactive MPD. The uncertainty about the other agent’s short- and long-term intent is represented using probabilistic effects of joint
actions. The focus of the thesis is on high-level, symbolic uncertainty in action recognition and preferences of the other agent, but may be extended to its low-level, geometric counterpart using existing methods in task and motion planning. Experiments are performed for different behavior models of the unknown agent, from the perspective of the protagonist or ego agent and are compared to baselines from sequential centralized planning. For most settings, the agents are able to eventually find the symbolic goal. Although the experiments are performed for a simple goal with two agents, the consistent convergence to the symbolic goal indicates that this approach may be extended to real-world settings with more complex ambiguity between actions and more agents acting collaboratively.

Aerial manipulation control is challenging due to the inherently coupled and highly variable dynamics involved in simultaneously controlling both the drone platform and its manipulator. Traditional model-based methods have been widely used but are limited by their dependence on accurate dynamics models, high computational cost, and sensitivity to external disturbances. Recent studies have explored reinforcement learning (RL)-based methods to overcome these limitations. However, most of them focus on fully actuated drones or rigid-link manipulators, thereby avoiding the complexity of underactuated platforms and manipulator control. This work develops a robust RL-based controller framework capable of whole-body control of an underactuated aerial manipulator with a two-degree-of-freedom arm.
The method is evaluated in both simulation and real-world experiments on three tasks: end-effector pose control, payload carrying, and path following. We achieve 6-DoF end-effector pose tracking with errors of 5.3 cm and 8.8°, with inference times of 0.18 ms, and demonstrate payload carrying capacity of up to 140 g through real-world experiments.

Autonomous navigation on inland waterways is challenging: channel geometry can vary widely, traffic rules are qualitative, and vessels must interact safely with other traffic while operating near infrastructure. This thesis presents a modular two-stage planning framework for a high-speed passenger ferry (Damen Waterbus) operating between Rotterdam and Dordrecht. A global planner extracts routes from Electronic Navigational Charts (ENCs) and refines them to produce a starboard-biased global path. A local planner uses Biased Model Predictive Path Integral (BIASED-MPPI) with ancillary controllers for path-following, goal reaching, and cruise speed regulation. Its cost function adapts based on the traffic scenario. Together, the planners generate starboard-biased trajectories that respect vessel limits, traffic and waterway constraints.

Global paths are built using the waterway-axis object extracted from the ENCs, and are made starboard-biased by projecting to the Fairway. This path is then refined using simple windowed smoothing schemes (minimum, average, and a hybrid), applied to the projection distances. The best smoothing configurations are explored via a grid search over various parameters. The generated paths are evaluated by metrics that assess feasibility. Across diverse intersections, smoothing noticeably reduces irregularity and sharp turns. The hybrid variant offers the most consistent gains, though feasibility in difficult geometries remains the dominant failure mode.

The local planner runs in real time and handles three encounter types (head-on, crossing, overtaking) using a standardised COLREG state definition to trigger rule-consistent behaviour. Variants without learned priors or with reduced replanning frequency are compared to highlight the effect of priors and update rate on success, stability, and computation time. Human-likeness is assessed against the Triton dataset, consisting of Waterbus AIS data via dock-to-dock distance and travel-time benchmarks.

Overall, the study demonstrates that combining ENC-informed global planning with a COLREG-aware MPPI controller yields safe, efficient, and human-like dock-to-dock trajectories in structured inland waterways, while identifying where the proposed method still limits performance.

We present Sadcher, a real-time task assignment framework for heterogeneous multi-robot teams that incorporates dynamic coalition formation and task precedence constraints. Sadcher is trained through Imitation Learning and combines graph attention and transformers to predict assignment rewards between robots and tasks. Based on the predicted rewards, a relaxed bipartite matching step generates high-quality schedules with feasibility guarantees. We explicitly model robot and task positions, task durations, and robots’ remaining processing times, enabling advanced temporal and spatial reasoning and generalization to environments with different spatiotemporal distributions compared to training. Trained on optimally solved small-scale instances, our method can scale to larger task sets and team sizes. Sadcher outperforms other learning-based and heuristic baselines on randomized, unseen prob-
lems for small and medium-sized teams with computation times suitable for real-time operation. Wealso explore sampling-based variants and evaluate scalability across robot and task counts. To address performance limitations identified for large-scale problem instances, we experiment with using Reinforcement Learning (RL) to fine-tune the imitation-learned model. Both discrete and continuous RL formulations are explored, leveraging Proximal Policy Optimization (PPO). This allows for training on larger problem instances, for which optimal solutions for IL are infeasible to obtain. Our RL experiments provide insights into the comparative advantages and trade-offs of IL and RL methods. In addition, we release our dataset of 250,000 optimal schedules to facilitate future research. We include a detailed description of the instance generation and the Mixed-Integer Linear Programming (MILP) formulation used to solve them optimally. Furthermore, this thesis contributes a lightweight
simulation environment with visualization tools for benchmarking different task assignment algorithms.

In the context of lunar exploration missions, autonomous navigation is essential to limit reliance on human intervention, which is hindered by significant communication delays and limited bandwidth. However, due to constrained on-board computational resources, complex robot-soil interactions, and the inherent challenges of full autonomy, most lunar rovers are currently limited to speeds of around 10cm/s. In this work, we investigate the use of Model Predictive Path Integral Control (MPPI) for high-speed navigation on uneven and cluttered lunar terrain. To this end, we develop an MPPI controller using the WARP framework, allowing to easily leverage the inherent parallelisation capabilities of GPUs. Experiments were conducted with a simulated Husky rover navigating realistic lunar terrain generated from NASA data, demonstrating the effectiveness of our motion planner. Since most MPPI implementations assume flat ground when sampling trajectories, this simplification can introduce inaccuracies, resulting in suboptimal, and sometimes even risky paths. To address this limitation, we propose a lightweight mathematical method for 3D trajectory projection, and evaluate its performance against the standard 2D MPPI approach. Overall, simulation results show that our 3D-enhanced MPPI significantly outperforms its 2D counterpart in terms of both traversal speed and obstacle avoidance on sloped, rock-scattered terrain. On flat ground, however, the traditional 2D MPPI still exhibits a high level of robustness. These findings suggest that an adaptive strategy, dynamically switching between 2D and 3D trajectory projection based on local terrain features, could offer a valuable trade-off between performance and power efficiency.

This paper presents the first decentralized method to enable real-world 6-DoF manipulation of a cable-suspended load using a team of Micro-Aerial Vehicles (MAVs). Our method leverages multi-agent reinforcement learning (MARL) to train an outer-loop control policy for each MAV. Unlike state-of-the-art controllers that utilize a centralized scheme, our policy does not require global states, inter-MAV communications, nor neighboring MAV information. Instead, agents communicate implicitly through load pose observations alone, which enables high scalability and flexibility. It also significantly reduces computing costs during inference time, enabling onboard deployment of the policy. In addition, we introduce a new action space design for the MAVs using linear acceleration and body rates. This choice, combined with a robust low-level controller, enables reliable sim-to-real transfer despite significant uncertainties caused by cable tension during dynamic 3D motion. We validate our method in various real-world experiments, including full-pose control under load model uncertainties, showing setpoint tracking performance comparable to the state-of-the-art centralized method. We also demonstrate cooperation amongst agents with heterogeneous control policies, and robustness to the complete in-flight loss of one MAV. Videos of experiments: https://autonomousrobots.nl/paper_websites/aerial-manipulation-marl

Autonomous planetary rovers require obstacle detection capabilities to navigate hazardous terrain without Earth-based intervention, yet currently deployed methods are limited. This research develops and validates a complete deep learning system for autonomous rock detection on a resource-constrained planetary rover.
We present a lightweight MobileNetV2-based U-Net architecture with dual attention mechanisms, optimized for edge deployment with only 0.31 million parameters. A new dataset MarsTanYard was created via a semi-automated dataset creation pipeline, enabling efficient annotation of Mars-analogue terrain imagery. Our deep learning network was trained on this dataset and integrated with a ROS2 navigation stack through a modular architecture that transforms segmentation masks into 2D occupancy grids that can be used for rover path planning. Our network achieves a 77% intersection-over-union accuracy for rock segmentation, and physical validation on a testing rover in a Mars analogue environment demonstrates a 94% detection rate for large rocks at close-range. An inference time of 4.49ms was achieved on the target rover hardware using model optimization techniques. The system maintains reliable operation across varying lighting conditions with less than 15% performance degradation.
Results show theoretical collision probability of 7.8 × 10^-7 per rock encounter, enabling months of autonomous operation for typical planetary missions. This work provides an end-to-end validation of the deep learning obstacle detection system, establishing a foundation for enhanced rover autonomy in future Mars exploration missions.

Autonomous UAV swarms have great potential for applications such as search‐and‐rescue, wildfire monitoring, and environmental mapping, but rapid and reliable coverage of large, obstacle‐filled areas remains challenging. In this paper, we ask: How can decentralised multi-UAV systems be optimised for large-scale priority-aware exploration? To answer this, we present the Fast LIDAR-based Autonomous Multi-UAV Explorer (FLAME), a hierarchical framework for efficient swarm exploration with long‐range, 360° LIDAR sensors. Locally, a dual‐mode exploration planner alternates between high‐speed straight‐line travel and frontier-cost‐based exploration. Globally, each UAV’s mission is optimised to visit user‐defined high‐priority regions early while minimising total mission time, enabling decentralised coordination with minimal duplicate exploration. FLAME also integrates battery‐constrained multi‐mission planning, dividing large tasks into multiple battery‐feasible missions with timely returns to the ground station for recharging or battery swaps. In simulation, FLAME achieves 36% lower mission time compared to state-of-the-art. A high-fidelity simulation confirms the framework’s ability to effectively prioritise user-assigned regions while adhering to operational constraints.

Autonomous drone racing presents a unique challenge that requires both high-speed motion planning and strategic decision-making in a multi-agent setting. Prior work has primarily relied on model predictive control (MPC) methods that treat opponents as dynamic obstacles, limiting their ability to model strategic interactions. In this work, we formulate drone racing as a dynamic game and introduce game-theoretic planning methods that compute open-loop Nash equilibria, incorporate blocking strategies, and accelerate decision-making using learning-based techniques. These methods explicitly model opponent behavior, allowing drones to anticipate and react strategically in high-speed racing scenarios. To assess the effectiveness of our approach, we conduct a large-scale head-to-head tournament against MPC-based planners, demonstrating that interaction-aware planning enables more effective overtaking and defensive strategies, leading to a higher wining rate. However, computational delays in high-speed decision-making can limit performance, highlighting the need for efficient techniques that balance real-time feasibility with strategic adaptability. Our results show that learning-based acceleration significantly improves decision-making speed while preserving competitive advantages. Finally, high-fidelity simulations and real-world drone racing experiments validate the feasibility of these methods, confirming their ability to generate reliable and competitive strategies under practical racing conditions.

Unmanned Surface Vessels (USVs) face significant control challenges due to uncertain environmental disturbances, such as waves and currents. This thesis proposes a trajectory tracking controller based on Active Disturbance Rejection Control (ADRC) implemented on the DUS V2500, a seagoing USV developed by Demcon Unmanned Systems. A custom simulation incorporating realistic wave, wind, and current disturbances is developed to validate the performance of the controller. Simulated experiments are supported by further validation through field tests in the harbour of Scheveningen, the Netherlands, and at sea. Simulation results demonstrate that ADRC significantly reduces cross-track error across all tested conditions compared to a baseline PID controller, but increases control effort and energy consumption. Field trials in Scheveningen, the Netherlands, confirm a reduction in cross-track error, while revealing a further increase in energy consumption during sea trials compared to the baseline controller. Modifications to the controller are proposed to improve its efficiency, and the impacts on tracking performance are discussed.

As robots increasingly operate in human environments, their controllers must ensure safe and reliable behavior under real-time constraints. Although optimization-based motion planners can enforce hard safety constraints, their computational demands limit their use on complex robotic platforms. Geometric motion planning offers a real-time alternative through optimization-free, closed-form control laws with reach–avoid guarantees. However, these guarantees rely on assumptions about obstacle representations that are often violated in realistic settings. When such assumptions fail, the planner’s dynamical system may preserve invariance of the safe set but lose global attractivity, jeopardizing goal reachability.

This thesis introduces a runtime verification algorithm, called Scenario-Shield, that adapts the geometric planner’s underlying dynamical system to expand its finite-time region of attraction. The method periodically samples nearby robot configurations and performs forward simulations to approximate this region. To accelerate this process, the approach is extended by incorporating statistical uncertainty quantification: conformal prediction is used to calibrate a fast membership test for candidate states, and the scenario approach provides a principled approximation of an uncountably infinite subset of the region of attraction.

To maintain computational efficiency, the algorithm is implemented using parallel computing and integrated into a geometric motion planning toolbox with ROS. The proposed method is validated in simulation on both a holonomic ground robot and a mobile manipulator, demonstrating improved reliability over baseline geometric fabrics controllers.

Open-world object manipulation has emerged as a popular research frontier in robotics. While recent advances in vision-language-action (VLA) models have achieved impressive results, they typically rely on large amounts of task-specific action data for training. This thesis aims to enable a manipulator to perform open-world object manipulation tasks without any action demonstrations. Instead of learning direct action mappings, we focus on understanding object dynamics.
To this end, we propose a novel framework that builds an explicit world model for open-world object manipulation. The framework integrates open-set segmentation and grasping, 3D digital twin reconstruction, and simulation-based strategy sampling within a unified framework. At the core of our approach lies the construction of a physically grounded digital twin of the environment, which enables the framework to simulate and evaluate diverse interaction strategies before real-world execution.
Experimentally, the proposed framework is able to perform multiple open-set manipulation tasks, such as “put the banana into the basket”, “stack the green cube onto the yellow cube”, and “place the blue cup upside on the wooden box”. These results are obtained without any task-specific action demonstrations, demonstrating strong generalization and autonomy compared to existing closed-set or imitation-based systems.

To safely and efficiently solve motion planning problems in multi-agent settings, most approaches attempt to solve a joint optimization that explicitly accounts for the responses triggered in other agents. This often results in solutions with an exponential computational complexity, making these methods intractable for complex scenarios with many agents. While sequential predict-and-plan approaches are more scalable, they tend to perform poorly in highly interactive environments. This paper proposes a method to improve the interactive capabilities of sequential predict-and-plan methods in multi-agent navigation problems by introducing predictability as an optimization objective. We interpret predictability through the use of general prediction models, by allowing agents to predict themselves and estimate how they align with these external predictions. We formally introduce this behavior through the free-energy of the system, which reduces (under appropriate bounds) to the Kullback-Leibler divergence between plan and prediction, and use this as a penalty for unpredictable trajectories. The proposed interpretation of predictability allows agents to more robustly leverage prediction models, and fosters a ‘soft social convention' that accelerates agreement on coordination strategies without the need of explicit high level control or communication. We show how this predictability-aware planning leads to lower-cost trajectories and reduces planning effort in a set of multi-robot problems, including autonomous driving experiments with human driver data, where we show that the benefits of considering predictability apply even when only the ego-agent uses this strategy.

Mobile manipulators, which integrate a robotic arm on a mobile base, are increasingly being explored and deployed in sectors such as healthcare, logistics, and aerospace. While motion planning for these systems has been studied in single-agent scenarios, the use of multiple robots to enhance efficiency and accelerate task completion in multi-agent settings remains largely unexplored, particularly in real-world environments. Extending motion planning to multi-mobile manipulators introduces challenges in real-time performance, collision avoidance, and coordination. To address these, this thesis proposes a decentralized Model Predictive Control (MPC) framework with a double integrator as dynamic model, denoted as MPC-d, tailored for multi-mobile manipulators operating in shared workspaces. It integrates optimization-based planning with robust state estimation, ensuring effective collision avoidance. Furthermore, a prioritized heuristic is introduced, leveraging the prediction horizon of MPC to resolve potential livelocks. The framework is validated through simulations and real-world experiments. Simulations compare MPC-d with MPC using a triple-integrator model (MPC-t) and a state-of-the-art geometric planner, called Geometric Fabrics (GF). Results demonstrate that MPC-d achieves comparable task success rates and collision avoidance compared to GF in pick-and-place scenarios while requiring less computation time than MPC-t. Real-world experiments confirm the framework’s viability, showcasing effective collision avoidance, enhanced efficiency from the prioritized heuristic, and consistency with simulation outcomes. Although MPC-d incurs higher computational costs than reactive geometric methods, it provides reliable performance and motion prediction of other agents in multi-agent settings.

In urban environments like Amsterdam, where road congestion and stress on infrastructure are critical issues, the city’s waterways remain underutilized for transportation. Autonomous Surface Vessels (ASVs), making waterway transport cheaper and less labor intensive, present a potential solution by reducing transportation costs and alleviating road traffic. Although companies are developing ASVs, current solutions are mainly limited to larger, less congested waterways, with further innovation needed for denser urban areas.

The challenges in autonomous navigation for ASVs in urban canals stem from the complex nonlinear dynamics of vessels, which require long-term planning and rapid responses to environmental changes. Urban waterways are narrow and unstructured, with loosely defined navigation rules that can lead to discontinuities in the motion planner’s cost function. Typically, motion planners rely on predictions of other agents’ movements and plan around them, resulting in no interaction awareness. This approach can lead to the freezing robot problem in dense environments, where the ASV halts, deeming all the space unsafe. A local motion planner must address these challenges by ensuring long-horizon interaction-aware motions, rule adherence, and real-time planning.

An emerging approach in autonomous navigation is a sampling-based Model Predictive Control (MPC) strategy known as Model Predictive Path Integral control (MPPI). This algorithm approximates the optimal control sequence by sampling from a continuous input distribution. Thanks to its gradient-free nature, MPPI only requires collision checking to avoid collision and allows discontinuous cost functions. Additionally, its computation speed remains largely unaffected by the complexity of the robots’ nonlinear dynamics, enabling longer planning horizons. While this approach has grown in popularity in the motion planning community, its application in dynamic environments with multiple interacting agents remains relatively unexplored.

Building on the state-of-the-art in MPPI, this thesis first introduces an Interaction-Aware Model Predictive Path Integral (IA-MPPI) controller tailored for motion planning in crowded urban canals. While conventional planners passively react to other agents, IA-MPPI actively predicts and plans cooperatively in real-time, addressing the freezing robot problem and handling discontinuous navigation rules. The decentralized, communication-free architecture assumes agent cooperation and employs a two-stage sample evaluation to enhance efficiency. This approach manages nonlinear dynamics, exact obstacle shapes, and longer planning horizons, demonstrating robustness and efficiency compared to state-of-the-art MPC in multi-agent environments.

While IA-MPPI uses a constant velocity model for predicting other agents’ hidden goals, Chapter 4 of this thesis introduces a learning-based trajectory prediction model trained on realistic artificial data to improve accuracy in crowded environments. By extracting the agents’ goals from predicted trajectories and integrating this information into the IA-MPPI controller, the planner’s performance increases significantly in environments with tight interactions. Moreover, this approach outperforms treating the predicted trajectory as occupied space, leading to safer navigation in dense urban canals.

One of MPPI’s main drawbacks is that it struggles in highly dynamic environments because it usually only samples around a previous plan. When rapid changes occur, such as strong disturbances or an obstacle unexpectedly cutting off the path, the previous plan becomes invalid, and all the generated samples may lead to collisions. To address this, Chapter 5 introduces Biased-MPPI, incorporating multiple classic and learning-based ancillary controllers into the sampling distribution. While the mathematical derivations show that this introduces a bias, this significantly improves reactivity, safety, and robustness to local minima. Additionally, this approach reduces the required samples, making the controller more efficient regarding computational demand.

Lastly, Chapter 6 leverages the gradient-free nature of MPPI and applies it to a different domain involving contact-rich manipulation tasks, which are notoriously difficult for model-based planning. With a parallelizable GPU-based physics engine like IsaacGym, MPPI can efficiently roll out hundreds of input sequences in parallel, faster than in real-time, to approximate optimal control. This approach eliminates the need for explicit modeling of contact dynamics by exploiting the models embedded in the simulator, making it particularly suited for manipulation tasks like pushing and picking with both prehensile and non-prehensile manipulators. Experiments with real robots demonstrate the effectiveness of this method in handling discontinuous, contact-heavy environments.

Overall, this thesis explores key aspects of motion planning, particularly for ASVs, with extensions to ground robots and mobile manipulators. It advances the MPPI framework for autonomous navigation by adapting it for multi-agent systems and introducing biases through classic and learning-based ancillary controllers. Additionally, the thesis compares MPPI to MPC in navigation experiments and to optimization fabrics in manipulation tasks. The work presented promotes the use of MPPI, especially in domains where complex nonlinear dynamics and discontinuous costs complicate the use of real-time optimization-based MPC.