Proactive Local-Minima-Free Robot Navigation: Blending Motion Prediction with Safe Control

TL;DR

This paper tackles safe, efficient navigation for autonomous mobile robots in dynamic, nonconvex environments with concave moving obstacles. It introduces an online barrier-learning pipeline that converts multimodal obstacle motion predictions into barrier functions via Gaussian Process Distance Fields and feeds them into an adaptive on-manifold Modulated CBF-QP (MCBF-QP) to avoid local minima. Two core contributions are the online prediction-to-barrier learning loop and an autonomous parameter-tuning scheme for deforming obstacle regions, enabling proactive, feasible navigation under uncertainty. The approach demonstrates improved safety and efficiency in simulations and real-world experiments, outperforming baselines in crowded, dynamic settings and enabling more robust obstacle negotiation than traditional reactive methods.

Abstract

This work addresses the challenge of safe and efficient mobile robot navigation in complex dynamic environments with concave moving obstacles. Reactive safe controllers like Control Barrier Functions (CBFs) design obstacle avoidance strategies based only on the current states of the obstacles, risking future collisions. To alleviate this problem, we use Gaussian processes to learn barrier functions online from multimodal motion predictions of obstacles generated by neural networks trained with energy-based learning. The learned barrier functions are then fed into quadratic programs using modulated CBFs (MCBFs), a local-minimum-free version of CBFs, to achieve safe and efficient navigation. The proposed framework makes two key contributions. First, it develops a prediction-to-barrier function online learning pipeline. Second, it introduces an autonomous parameter tuning algorithm that adapts MCBFs to deforming, prediction-based barrier functions. The framework is evaluated in both simulations and real-world experiments, consistently outperforming baselines and demonstrating superior safety and efficiency in crowded dynamic environments.

Figures (5)

• Figure 1: General pipeline of the proposed MMP-MCBF approach. Given a perception and tracking system, a motion predictor makes predictions of dynamic obstacles, which are converted into barrier functions via online learning using Gaussian Process Distance Fields (GPDFs). The proposed adaptive MCBF controller utilizes the barrier functions to generate safe and efficient control actions. In this work, the learning-based motion predictor is pre-trained.
• Figure 2: Modulated CBF (left) vs. CBF (right). When facing a complex concave obstacle (in purple), modulated CBFs generate an additional exit "force" (long red arrows) introduced by Eq. \ref{['eq:mod_phi_cbf_constraint_affine']}, guiding the robot along the isoline (dashed lines) direction $\phi(\bm{x}, h_o)$ and away from the obstacle, whereas standard CBFs become trapped in local minima.
• Figure 3: Performance of tangent vector selection using geodesic approximation with constant $\beta$ (first row until the dashed divider, $\beta=0.02$ in 2D, $\beta=0.05$ in 3D) versus generalized geodesic approximation with adaptive $\beta$ selection (second row until the dashed divider). Pink polygons are the estimated forward reachable space of the three humans from the motion predictor under fixed 0.05s time steps and 60 iterations. The last column is the visualization of the augmented CBFs used in the 3D cases.
• Figure 4: Hospital (left, middle) and crowd (right) simulation environment setup for robot comparison tests, including the Gazebo world (left) and the static environmental layout in the robot's perspective (middle, right).
• Figure 5: Real-world experiments. A Fetch robot avoids three pedestrians approaching in a U-shape formation. Full video attached.