In this paper, we investigate a hybrid control paradigm that merges nonlinear model predictive control (MPC) and model-based reinforcement learning (RL) for autonomous navigation of a model car across off-road, unstructured terrain without predefined maps. Inspired by BADGR, an LSTM-based network that predicts future events from an image and a sequence of control actions, we investigate the replacement of LSTM modules with transformers to significantly improve predictive performance. To address uncertainty, we train an ensemble of models that estimate the mutual information between weights and outputs, facilitating dynamic horizon planning through variable speeds. To account for the intricacies of our vehicle's model and states, we also incorporate a nonlinear MPC into our control loop. The model-based RL module produces steering angles and quantifies uncertainty, while the nonlinear MPC suggests optimal throttle settings. This allows us to strike a balance between goal attainment speed and model uncertainty. Our approach consistently outperforms existing baselines in predicting future events and integrates the vehicle's kinematic model for enhanced decision-making.

In this paper, we address the challenge of underwater object tracking using a two-part approach: perception and control. Our method leverages AQUA, a highly maneuverable autonomous underwater vehicle (AUV), for open-water experiments. We employ PID controllers for AUV control and a spiral search algorithm for target recovery. Our approach relies solely on observations in the image plane, eliminating the need for robot localization or camera calibration. By combining YOLOv7 for target detection, SORT filtering for temporal stability, and the spiral search algorithm for target recovery, we demonstrate strong long-term tracking performance across both offline datasets and open-water experiments in harsh ocean conditions in Barbados.

Due to intricate interactions between the ocean and the atmosphere, predicting the paths of nonactuated floating objects in the ocean is challenging, especially over long-term horizons. Despite these complexities, long-term trajectory prediction is vital for search and rescue operations, ecological research, and disaster response. Inspired by the DARPA Forecasting Floats in Turbulence challenge, we introduce an open-source benchmark dataset for evaluating oceanic trajectory prediction models. This dataset combines ocean drifter paths and historical wind and current data, enabling the development of machine learning (ML) models tailored to the complexities of drifter trajectory prediction. Alongside this benchmark dataset, we offer a baseline solution set created using OpenDrift, an open-source tool for modeling object trajectories in the ocean and atmosphere.