In-context algorithm distillation-based maneuvering evasion scheme for high-speed flight vehicles under limited measurement information

This paper addresses the challenge of intelligent evasion for high-speed flight vehicles under limited measurements (only partial observations such as line-of-sight angles). This restriction transforms the problem into a Partially Observable Markov Decision Process (POMDP), where conventional reinforcement learning policies struggle to generalize to unseen scenarios. This paper proposes a novel meta-learning solution based on Algorithm Distillation (AD). The resulting AD model exhibits in-context learning, enabling zero-shot generalization to new threat scenarios without gradient updates. Simulations show the AD method achieves substantially higher success rates across diverse test scenarios than both the original PPO and an expert-distillation baseline, narrowing the gap between static RL policies and the demands of real-world evasion tasks.