Safe Real-Time Obstacle Detection and Navigation Using CBF–CLF and CBF–PID Control

Traditional robotic navigation pipelines typically follow a three stage architecture: obstacle detection, path planning, and low-level control for trajectory tracking. While effective in static environments, these methods often introduce latency and lack formal guarantees of safety in dynamic or unplanned for scenarios. Our work addresses these limitations by developing a real-time controller grounded in Control Barrier Functions (CBFs) and Control Lyapunov Functions (CLFs), unified through a Quadratic Program (QP). We first investigate a hybrid CBF-PID-QP controller on a 1/10 scale car, where the CBF serves as a real-time safety filter, modifying the PID output to prevent constraint violations. While this architecture guarantees safety, it can result in “sticking points” or equilibrium states where the control inputs satisfy safety constraints but fail to make progress toward the goal. This is particularly pronounced under unicycle dynamics, where steering inputs may not influence the CBF, making certain maneuvers infeasible. To overcome this, we transition to a full CLF-CBF-QP framework using single-integrator dynamics. This formulation eliminates the need for global path planning or prior exploration. The CLF ensures convergence to a goal position, while the CBF maintains safety, allowing the controller to generate control inputs that are both safe and goal oriented in real time. We further propose a novel modification to the CBF function that enables prioritization of obstacles directly in the robot’s path and vii ensures a non-vanishing gradient, effectively eliminating sticking points even in symmetric obstacle-goal configurations. This method is validated on the Unitree G1 humanoid. Across simulation and hardware experiments, we demonstrate that our formulation avoids sticking points and successfully handles both static and dynamic obstacles without relying on global planners or learning- based policies. This work highlights that real-time, low-level controllers can replace conventional planning stacks, providing safe, responsive, and mathematically grounded behavior in complex, dynamic environments.