Reward inference from expert demonstrations and tail-risk-sensitive sampling control for pursuit-evasion games of underactuated ASVs

Adversarial marine missions place underactuated autonomous surface vehicles (ASVs) in pursuit-evasion games (PEGs) in which nonholonomic actuation, imperfectly known hydrodynamics, and mutually reactive opponents are encountered simultaneously. A demonstration-driven control architecture is developed, in which the optimisation objective itself, rather than the controller alone, is recovered from data. Behavioural features spanning target proximity, obstacle clearance, speed utilisation, capture likelihood, and inter-agent separation are encoded by radial basis expansions, and a lightweight inverse reinforcement learning (IRL) scheme recovers the associated weights by matching the feature expectations of expert trajectories under a maximum-entropy model. Reusable dynamics priors are supplied by a three-phase transfer pipeline: state spaces of unequal dimension are accommodated by a sparsely connected weight-sharing backbone, the adaptation sequence is reorganised by a difficulty-ordered curriculum, and the adapted model is compressed by teacher-guided distillation to a latency budget compatible with onboard execution. Online commands are produced by a tail-risk-sensitive model predictive path integral (MPPI) scheme, in which rollout scoring is driven by the inferred reward, unfavourable outcomes are penalised through the conditional value-at-risk (CVaR) of the rollout disutility evaluated separately for every candidate control sequence, the exploration–exploitation balance is retuned by a range-modulated temperature as the engagement tightens, and the proposal distribution is biased towards admissible manoeuvres by the transferred dynamics. Discrete target assignment is delegated to an upper-level allocator re-solved from the real-time states, so that team coordination is separated from continuous command synthesis. Hardware-in-the-loop experimental results verify the effectiveness of the proposed method. Across 24 paired HIL cases, the proposed method achieves 100% capture success and a median capture time of 29.75 s, compared with 79.2%–95.8% and 31.85–38.75 s for three baselines; its median P95 controller computation time of 3.011 ms remains below the 0.1 s control period.