Meta-Learning Enhanced Risk-Aware Adaptive Motion Planning in Unknown Environments

Safe and efficient motion planning in unknown, cluttered environments is a critical challenge for autonomous robots. This paper introduces a meta-learning-enhanced motion planning framework that integrates multi-scale risk prediction with adaptive model predictive control (MPC) to address this challenge. In this work, we propose a unified meta-learned model for signed distance function prediction, termed Meta-SDF-a single neural network that learns generic geometric priors from diverse obstacle types during offline training. At runtime, the Meta-SDF rapidly adapts to novel, unseen obstacles using only a small amount of point cloud data, without prior knowledge of obstacle geometry. For online planning, we develop a risk-aware adaptive MPC framework augmented with control barrier functions (CBFs), where safety constraints are derived from the adapted Meta-SDF. A key innovation is a multi-scale risk prediction mechanism that evaluates collision risk across short-, mid-, and long-term horizons within the MPC prediction window. This risk metric dynamically guides an adaptive strategy that adjusts both the MPC solving frequency and prediction horizon: high-risk scenarios trigger high-frequency solving with a shortened horizon for reactive obstacle avoidance, while low-risk scenarios employ lower-frequency solving with an extended horizon for efficient goal-directed planning. Simulation experiments demonstrate the efficacy of the proposed framework.