DPI-TD3: Data-Driven Evasive Maneuver Strategy for Adaptive Control of Exo-Atmospheric Vehicles

In the context of evasive maneuvering for exo-atmospheric vehicles, reinforcement learning and other data-driven decision-making techniques have been explored extensively. However, most existing studies focus on scenarios where interceptors use a single guidance strategy, leading to significant performance degradation when the vehicle faces interceptors with different strategies. To address this, we introduce a novel Deep Policy Inference Twin Delayed Deep Deterministic Policy Gradient (DPI-TD3) algorithm that enhances evasive capabilities against interceptors employing a variety of guidance laws. We present a interception simulation framework that includes multiple types of interceptors. The deep policy inference model identifies the guidance law of the interceptor using the relative motion vector between the interceptor and the vehicle. Depending on the identified interceptor type, the algorithm either reuses an existing experience buffer or creates new ones through deep Bayesian inference and an experience mixing network. The updated TD3 algorithm then uses the selected buffer to train against the current interceptor, generating acceleration commands for the vehicle. Experimental results show that, compared to baseline methods, the proposed algorithm converges faster and produces more effective evasive maneuvers in response to various guidance laws. Under baseline conditions, DPI-TD3 achieves a penetration success rate of 96.4% and a miss distance of 15.58 m, outperforming TD3, Deep Deterministic Policy Gradient (DDPG), and the differential game method. In more complex scenarios with sensor noise and reduced interceptor maneuverability, DPI-TD3 still maintains success rates of 92.5% and 92.3%, showing less performance degradation than baseline methods.