Real-Time Data-Driven Safety-Critical Control in Unknown Multiobstacle Environments

Real-time obstacle avoidance in unknown dynamic multiobstacle environments is critical for autonomous robotic systems, particularly when relying solely on onboard LiDAR sensors without prior environmental knowledge. Control barrier functions (CBFs) provide a theoretical foundation for ensuring safety in such scenarios. However, existing learning-based CBF approaches still face three key challenges: LiDAR-based methods suffer from prediction inconsistencies due to sensor noise between consecutive time steps; current learning methods require offline training or prior environmental information or exhibit poor real-time performance below 10 Hz in multiobstacle environments; and existing approaches cannot synthesize velocity-aware terms essential for dynamic CBFs (D-CBFs) to handle moving obstacles. To address these limitations, this article introduces a velocity-aware fast adaptive variational sparse Gaussian process (VA-FAVSGP) safety framework that constructs a class of buffer-robust D-CBFs utilizing only online LiDAR data. The framework makes three key contributions: First, to overcome prediction inconsistencies caused by sensor noise and improve real-time performance, we develop a VA-FAVSGP model with historical data integration and temporal weighting mechanisms for CBF learning that ensures smooth and consistent CBF predictions, achieving 70–90 Hz (slow mode) and 320–420 Hz (fast mode) incremental learning without offline training requirements. Second, to enable real-time dynamic obstacle handling, we exploit the translation invariance property of radial basis function kernel functions to construct velocity-aware terms for D-CBFs. Third, we further construct learning-based buffer-robust D-CBFs with selective activation strategies that maintain computational efficiency regardless of obstacle quantity. Theoretical analysis proves system safety and asymptotic stabilization. Extensive experiments demonstrate that the proposed framework achieves reactive frequencies of 25–40 Hz (slow mode) and 50–65 Hz (fast mode) even when integrated with multistep predictive control, efficiently handles static, sudden-appearance, and dynamic obstacles, and exhibits better performance compared to state-of-the-art methods in three-dimensional environments with 2-D LiDAR-based perception.