Khalasi: Energy-Efficient Navigation for Surface Vehicles in Vortical Flow Fields

TL;DR

Khalasi addresses energy-efficient autonomous surface navigation in vortical flows under partial observability by combining Gaussian Process Regression-based flow reconstruction, CNN-encoded local flow features, and Soft Actor-Critic learning. Trained across diverse simulated flow environments, it demonstrates substantial energy savings (30–50%) and strong generalization to unseen flow conditions and NOAA Currents without retraining. The approach relies solely on local velocity measurements and a lightweight observation stack to infer flow structure, enabling robust long-duration autonomy. The work provides a scalable benchmark and open-source resources to advance research on flow-aware navigation for ASVs.

Abstract

For centuries, khalasi (Gujarati for sailor) have skillfully harnessed ocean currents to navigate vast waters with minimal effort. Emulating this intuition in autonomous systems remains a significant challenge, particularly for Autonomous Surface Vehicles tasked with long duration missions under strict energy budgets. In this work, we present a learning-based approach for energy-efficient surface vehicle navigation in vortical flow fields, where partial observability often undermines traditional path-planning methods. We present an end to end reinforcement learning framework based on Soft Actor Critic that learns flow-aware navigation policies using only local velocity measurements. Through extensive evaluation across diverse and dynamically rich scenarios, our method demonstrates substantial energy savings and robust generalization to previously unseen flow conditions, offering a promising path toward long term autonomy in ocean environments. The navigation paths generated by our proposed approach show an improvement in energy conservation 30 to 50 percent compared to the existing state of the art techniques.

Figures (8)

• Figure 1: Overview of khalasi pipeline. (A) Training and evaluation environment: The agent (red) navigates toward the goal (orange) by leveraging local flow features. The green box and orange box denotes the agent and goals spawn region respectively. (B) Training pipeline: historical local velocity observations are processed through a Gaussian Process Regression (GPR) module for flow reconstruction, followed by spatial gradient extraction via a CNN encoder. The resulting latent representation, combined with positional and goal data, is input to a Soft Actor–Critic (SAC) policy. (C) Trained using parallelized environments with diverse flow variations enable better generalization across different flows.
• Figure 2: Example of flow reconstruction in a vortex environment using GPR. Accuracy improves as more temporal samples are incorporated, shown by predictions at $t$, $t+10$, and $t+20$ compared to the ground truth.
• Figure 3: Mean Absolute Error (MAE) of GPR-based flow reconstruction compared to ground truth, evaluated along a dummy agent trajectory in the same environment.
• Figure 4: (A) Environment setups shown from left to right: oscillating single cylinder, static single cylinder, and static double cylinder. (B,C) Spawn regions used for training and testing: vertical spawn and L-shaped spawn, respectively.
• Figure 5: Sample agent trajectories in different flow environments. The color gradient along the path represents the energy used by the agent at each point. Note: the background flow shown in the figures corresponds to the initial flow at the start of each trial. (A) Trajectories in the oscillating single-cylinder environment. (B) Trajectories in the static single-cylinder environment (normal von Kármán vortex street). (C) Trajectories in the static double-cylinder environment.