This thesis presents a comparative analysis of multi-target tracking algorithms for estimating the kinematic state and physical extent of maritime vessels from 4D millimetre-wave radar point clouds. The research addresses a gap in understanding the trade-offs between algorithmic complexity and motion model fidelity for inland maritime navigation.
The study evaluates three tracking paradigms, Tracking-by-Detection (TBD), the Probability Hypothesis Density (PHD) filter, and the Poisson Multi-Bernoulli Mixture (PMBM) filter, each tested with linear (Constant Velocity, Constant Acceleration) and non-linear (Constant Turn Rate) motion models. To represent vessel shape, three extended object models were implemented: the Random Matrix, Star-Convex Gaussian Process, and Principal Axis models.
Performance was evaluated using a mix of simulated datasets with perfect ground truth and collected real-world datasets. A hyperparameter optimisation process was conducted for every algorithm combination using the Optuna framework, with the objective of minimising the Generalised Optimal Sub-Pattern Assignment metric using a specified cost function.
The results yielded several key, and at times counterintuitive, findings. Firstly, increased algorithmic complexity did not yield superior performance. The simpler TBD and PHD frameworks consistently outperformed the theoretically optimal but computationally intensive PMBM filter, which was prone to generating unstable tracks and excessive false positives. Secondly, non-linear motion models offered no significant advantage over the simpler linear models. This could be attributed to the slow dynamics of maritime vessels, the high sensor update rate, which makes linear extrapolation sufficient, and the fact that measurement noise often dominates any subtle gains from a more precise motion model. Thirdly, for extent estimation, the Random Matrix model demonstrated the best balance of accuracy, stability, and computational efficiency. The more complex Gaussian Process and Principal Axis models struggled with stability and practicality.
The study concludes that the optimal balance is achieved by pairing simple linear motion models with the Random Matrix extent model, yielding accurate cardinality estimation, reliable state tracking, and efficient computation. The findings underscore that robustness and interpretability outweigh algorithmic sophistication when designing maritime tracking systems based on 4D radar. Ultimately, the thesis establishes a principled performance baseline and highlights the diminishing returns of complexity in this domain.

Through the use of Eigenmanifold theory, unforced periodic trajectories called nonlinear normal modes can be identified and excited in nonlinear mechanical systems. Applying this to repetitive tasks in robotic systems with elastic components can drastically reduce energy consumption, as normal modes in steady-state require no additional control effort. However, existing control methods for exciting nonlinear normal modes have so far only assumed full actuation. Consequently, these techniques are incompatible with series elastic joint robots, even though they represent a significant subclass of physical systems with elastic elements. Additionally, the calculation and parameterization of Eigenmanifolds for high-dimensional systems generally remains a complex task and is difficult to scale. While existing literature aims to avoid forced evolutions or model cancellation, we instead lean into this approach. By rephrasing Eigenmanifold-based control as a trajectory tracking problem, standard techniques for elastic joint robot trajectory tracking control can be employed. Furthermore, obtaining theoretical guarantees on global stability becomes possible. In this work, a new modular control architecture is presented that integrates trajectory tracking feedback control with Eigenmanifold theory to dynamically generate and track hyper energy-efficient oscillatory movements in underactuated systems. This approach enables the excitation of nonlinear normal modes using standard trajectory tracking controllers, while preserving energy-efficient properties desired from Eigenmanifold-based controllers. We first discuss the theoretical validity and energy-efficiency of this control architecture, and then test the architecture in simulation for a variety of use-cases and controllers.

Autonomous mobile robots, once confined to structured settings like warehouses, now operate in dynamic, human-centered environments such as hospitals and streets, where interactions with humans are unavoidable. This shift poses major challenges for local motion planning, requiring robots to navigate complex environments where human behavior is difficult to predict while prioritizing safety, adaptability, and social compliance. This thesis addresses these challenges by proposing solutions for interaction-aware, safe, socially compliant, and adaptive planning.

We first examine the standard approach of splitting local motion planning into two sub-problems: predicting human trajectories and solving constrained trajectory optimization to avoid collisions. While this approach provides safety guarantees and can handle uncertainties, it neglects the interactions between the robot and the humans. We address this by formulating a Model Predictive Control problem in which the robot’s action influences both its own state and the human states. Focusing on navigation among pedestrians, we leverage the interpretable and established Social Force Model (SFM) to model the human response dynamics to robot actions. By accounting for the robot’s influence on pedestrian behavior, we demonstrate that the robot can guide pedestrian behavior. However, a carefully designed cost function is essential to promote cooperation and prevent exploitation.

Next, we focus on the limitations of learning-based approaches that can address social compliance, specifically, Imitation Learning (IL). While IL enables the learning of socially compliant behaviors from demonstrations or observations, the resulting policy lacks formal safety guarantees. Safety filters based on, e.g., Control Barrier Functions (CBFs), adapt control inputs to ensure safety and can be combined with IL policies. While CBFs are an effective tool to certify safety, two challenges remain: constructing them for complex systems with input constraints and accounting for model uncertainties. We propose Robust Policy Control Barrier Functions, a method for constructing robust CBFs that guarantees safety under worst-case bounded disturbances. Furthermore, we present a practical approximation and demonstrate its effectiveness in simulation and hardware experiments.

Next, we focus on adaptability, specifically in data-driven pedestrian prediction models, which are crucial for the decoupled prediction and planning approach. While existing models are trained offline on general datasets, they may not reflect the behavior of pedestrians in the robot’s environment. To address this, we propose a self-supervised continual learning framework that refines models during deployment using online data from the robot’s perception pipeline, preserving prior knowledge through regularization and selective retraining. Experiments show improved performance compared to naive online training. Although we focus on pedestrian prediction, the approach could extend to other methods applying learning-from-observations.

In summary, this thesis integrates interaction awareness into decoupled motion planning by leveraging the established SFM, presents a method for constructing safety filters using practical approximations of robust CBFs, and develops a framework that addresses adaptability through self-supervised continual learning. Integrating the safety filter with a socially compliant, interaction-aware policy learned from observations and adapted online through continual learning offers a comprehensive solution.
This combination could pave the way for local motion planning that is interaction-aware, safe, socially compliant, and adaptive.

Spatiotemporal systems are systems whose dynamics depend on time and space and are commonly found in real life. These systems are mathematically modeled using partial differential equations and are also known as distributed-parameter systems. Due to their structure and the high number of variables involved, control and estimation for this class of systems are very challenging. This thesis addresses two problems related to spatiotemporal systems: state estimation and system identification.

Monitoring the states of a control system is important to ensure the behavior of the system is achieving the control objectives. This can be achieved, among others, by using state observers that estimate the states of the systems regularly. First, we present a literature review of observer design methods for distributed parameter systems. In general, the design requires a dimension-reduction approach to implement the observer. From the dimension reduction, the design approaches can be classified into late and early lumping. In the late lumping perspective, model reduction is performed at the end of the observer design. In the early lumping perspective, dimension reduction is applied to the model of the system. We incorporate both approaches in our literature review.

State observer design requires the model of the systems. This thesis also presents a system identification method for distributed-parameter systems. The identification of such systems typically requires spatially dense and regular measurements, followed by selecting sensors that provide significant measurements to the model to reduce the model complexity. However, these requirements may be challenging to fulfill. In case the sensor locations are irregular and sparse in space, we propose the use of lumped-parameter system identification.

For models with a large number of regressors, we propose a method for reducing the number of regressors using a tree representation. The tree is a way to list models with different numbers of regressors. From all possible regressors for the model, the proposed method builds the tree from the simplest models, i.e., models with one regressor. The number of regressors in the models is incrementally increased to one or more models with the best performance. The addition is repeated until the tree contains models with the desired maximum number of regressors.

System identification is typically performed using a complete data set, i.e., for each input sample, there is an associated output sample available. However, there are cases in which some output samples are not recorded in the data set, making the identification data incomplete. This thesis also considers the problem of incomplete data for Takagi-Sugeno (TS) fuzzy system identification using the product space clustering method. This method comprises two steps: fuzzy clustering and rules construction. The first proposed method enables the use of incomplete system identification data to fuzzy c-means clustering algorithm developed for incomplete classification, which yields different estimates for a missing sample. This can be achieved by fusing those different values into a single value. The second proposed method treats missing samples as optimization variables during the identification process. The optimization is repeated until the change of all optimization variables is small.

Reinforcement learning has emerged as a promising approach for enabling robots to learn from interactions with their environments, without relying on predefined behaviors. However, robots face significant challenges when learning directly from real-world interactions. Real-world learning is time-consuming and resource-intensive, often requiring extensive data collection over long periods. Additionally, the risks involved in trial-and-error learning in physical settings are high, as faulty policies can lead to safety issues or system damage. Simulations offer a safer and more efficient alternative, allowing robots to learn in simulated environments at faster-than-real-time speeds. Despite these benefits, simulations often serve as imperfect approximations of reality. As a result, robots may learn behaviors that exploit simulation-specific quirks, which may not perform well in real-world settings, creating difficulties in transferring learned behaviors from simulation to real environments—a challenge known as the sim-to-real gap. Several factors contribute to the sim-to-real gap, such as unmodeled physical phenomena like friction and deformation, and the asynchronous nature of real-world systems that simulations often fail to capture accurately. Additionally, using separate software stacks for simulation and deployment can unintentionally lead to discrepancies. Finally, simulating at faster-than-real-time speeds with asynchronous frameworks that distribute computation across multiple cores may also introduce inaccuracies without proper synchronization.

This thesis focuses on improving simulation tools and methodologies to enhance the efficiency and effectiveness of learning-based approaches in robotics. The work addresses key trade-offs between flexibility, speed, and accuracy in robotic simulations, which are critical for successfully transferring learned policies from simulation to real-world environments. Additionally, it introduces a strategy to improve resilience, ensuring that learned behaviors are robust to irrelevant and unknown dynamics.
By tackling these challenges, this thesis provides insights into the design of effective robotic simulators and presents contributions that help bridge the gap between simulated and real-world robotic learning.

As we increasingly strive to integrate robots into human-centric environments, safety is a top priority. Traditionally, rigid collaborative robots have relied on safety-aware computational control policies, which are susceptible to perception errors and often lead to overly cautious behavior that limits performance. In contrast, soft robotics offers a promising alternative by ensuring passive compliance throughout the robot’s structure via material softness. This mechanical compliance inherently mitigates safety issues arising from perception or control errors, although this has been paid with a substantial drop in precision. In recent years, significant advances have been seen in soft robotics, with exciting new developments in design, smart materials, actuators, sensors, models, and control strategies. However, the modeling and control of continuum soft robots continue to pose major challenges due to their infinite degrees of freedom, complex nonlinear dynamics, and time-dependent behaviors like hysteresis. As a result, soft robots often lack the necessary capability and motion precision, leading to a tradeoff where performance is sacrificed for safety. With this thesis, we argue that this tradeoff can be overcome by developing more advanced algorithms that can reason on the physics of the soft robot. More specifically, we propose combining powerful learned models with efficient and effective model-based control approaches that allow for interpretability into the actions and admit stability guarantees.

Currently, two main approaches exist for controlling soft robots. The first employs model-based control using approximated physics-based models derived from first principles. The second directly learns control policies, primarily through reinforcement learning. Both strategies face notable limitations. Existing model-based controllers are unable to fully manage and eventually exploit the dynamics of soft robots because their underlying models inadequately capture complex behaviors, particularly how actuation and external interactions affect the robot’s deformation. Moreover, deriving these models requires extensive expert knowledge. Additionally, the combined complexity and uncertainty of the dynamics between a soft robot and its environment make it currently infeasible to develop comprehensive world models from first principles alone, thereby motivating the integration of machine learning approaches that can effectively leverage data-driven insights. Conversely, directly learning the controller — such as via reinforcement learning — lacks interpretability and stability guarantees while being highly sample inefficient, a significant drawback given the time-dependent material properties and limited lifespan of current soft robots.

In this thesis, we contend that combining learned models with model-based controllers presents a promising alternative that brings together the advantages of both approaches: expressive, data-driven models that require less expert knowledge paired with controllers that are both interpretable and provably stable. Although recent years have seen increased interest in leveraging learned models for control, most work in this area depends on computationally intensive optimal control methods, such as MPC, to optimize the actuation sequence with the learned model. However, the high computational cost of solving these optimal control problems limits the maximum control frequency during deployment, preventing us from fully exploiting the dynamic capabilities of soft robots. Instead, this thesis pursues closed-form controllers that utilize the physical structure of learned models within an energy-shaping framework. The main challenge here is to develop approaches that integrate such physical structures—specifically, kinetic and potential energy terms—into the learning of dynamical models for soft robots. Before addressing this main challenge, we first had to advance physics-based models derived from first principles and identify novel techniques to leverage them for control. On one hand, this clarified which physical priors were available for learning, while on the other hand, it inspired new ways to integrate model-based controllers with learned models. The thesis addresses this topic through several interconnected key contributions.

First, we argue that quantifying the safety of soft robots is crucial for designing and controlling them to ensure that the closed-loop system meets the specific safety requirements of their intended applications. To this end, we present the first safety metric for continuum soft robots, which assesses the safety of an integrated soft robot design by accounting for both its embodied and computational intelligence.

Secondly, this thesis enhances shape sensing for soft robots by leveraging insights from kinematic models. We accomplish this by formulating and solving nonlinear optimization problems that align sensor measurements with the backbone shapes predicted by the kinematic model. We present two distinct approaches that integrate commercial sensors—namely visual and magnetic—with SLAM algorithms and a learned sensor measurement model, respectively, to accurately estimate the soft robot’s state, a key requirement for effective feedback control.

Thirdly, this thesis introduces advanced physics-based actuation models, including those for robots actuated by auxetic metamaterials - referred to as HSA robots—and models that capture the actuation dynamics of piston-driven pneumatic soft robots. We then leverage the acquired model insights to design provably stable nonlinear controllers—specifically, an integral-saturated PID combined with potential shaping and Cartesian-space impedance control for planar GSA robots, as well as a backstepping controller for pneumatic piston-driven soft robots. This contribution deepens our understanding of actuation, a critical aspect of soft robot behavior, and demonstrates how such insights can be incorporated into model-based control strategies. Moreover, experiments with HSA robots have highlighted the limitations of purely physics-based models in capturing complex phenomena like hysteresis, thereby motivating the exploration of learning-based approaches. In the future, the developed actuation models can serve as valuable physical priors for learned models.

Fourthly, the thesis presents techniques for learning soft robot models that incorporate physical structures while ensuring stability. We accomplish this by embedding physics-based dynamical models into the learning algorithm, which determines the free parameters of the dynamics and optionally optimizes a coordinate transformation—such as encoding into latent space. Two notable approaches are introduced: (1) an algorithm that extracts low-dimensional soft robot strain models from samples of the robot backbone’s shape evolution, and (2) a network of coupled harmonic oscillators for learning latent dynamics from high-dimensional observations like images. The explicit inclusion of kinematic and potential energy terms in these models allows for stability analysis using standard nonlinear system theory tools, such as Lyapunov methods. For instance, we prove that the coupled oscillator network is both globally asymptotically stable and input-to-state stable.

Fifthly, we exploit the physical structure of the learned models from contribution four to design closed-form setpoint regulators. The controller contains two key components: (1) a potential shaping feedforward term that positions the local/global minimum of the closed-loop potential energy at the setpoint by leveraging the learned model knowledge, and (2) an integral-saturated PID feedback term that rejects disturbances and compensates for modeling errors to prevent steady-state errors. The stability of the closed-loop system can then be analyzed using Lyapunov arguments.

Finally, the thesis explores methods for generating compliant motion behaviors in soft robots beyond low-level control. One approach focuses on assisting users, particularly elderly individuals, with activities of daily living by guiding the low-level controller with brain signals. This is achieved by combining motor imagery classification from wearable EEG devices with compliant impedance control in operational space. The second approach combines an orbitally stable dynamical system in latent space with a bijective neural network parametrized encoder to learn periodic motions from demonstrations. By avoiding reliance on time references, this learned motion policy enables natural and compliant tracking of demonstrated periodic motions. This contribution ensures that not just the robot structure and low-level controller are compliant, but also the high-level motion strategy.

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.