Active visual SLAM using potential fields and model predictive control

• Integrated Active Visual SLAM combining potential fields and MPC. • Online MPC for collision-free, constraint-compliant, and utility-optimized motion. • Real-world experiments with a non-holonomic robot. • Successful autonomous exploration of unknown environment. • Reactive adaptation to dynamic environments. Active Simultaneous Localization and Mapping (SLAM) enables robots to autonomously explore unknown environments while building a map and localizing themselves. In indoor construction sites, this capability can support efficient and safe exploration as a basis for subsequent automated construction processes. This paper presents an integrated Active Visual SLAM approach combining potential fields with Model Predictive Control (MPC). The framework builds on ORB-SLAM3, extended with a trajectory generation module that first produces a dense 2D Probabilistic Occupancy Grid Map. A behavioral exploration strategy then employs potential fields modeled as multivariate Gaussian distributions to attract the robot toward unexplored and frontier regions while repelling it from occupied areas. A discrete, nonlinear multi-stage MPC generates feasible trajectories by optimizing the robot’s trajectory over the global potential field online, ensuring compliance with specified optimization goals, and enforcing collision avoidance as well as non-holonomic constraints. Experiments with a six-wheel skid-steering robot in a test environment resembling indoor construction sites demonstrate effective exploration and robustness to dynamic changes, validating the approach in real-world scenarios.