Synergistic Hierarchical AI Framework for USV Navigation: Closing the Loop Between Swin-Transformer Perception, T-ASTAR Planning, and Energy-Aware TD3 Control

. 2025 Jul 30;25(15):4699. doi: 10.3390/s25154699

Algorithm 1 T-ASTAR: Transformer-driven Adaptive Search and Reasoning

Input: Start

v_{s t a r t}

, Target set of states

V_{g o a l}

, Environmental observation

O

, Maximum number of iterations

K_{m a x}

Output: Optimal path

π^{*}

, Path cost

C_{b e s t}

1
$E = T r a n s f o r m e r E n c o d e r (O); \forall v \in v_{s u v_{g}} : h_{v} = E m b e d (v, E)$
2
$V_{u} = \{v_{s}\}; Q_{v} = \{v_{s}\}; Q_{E} = \emptyset; C^{*} \leftarrow \infty; π^{*} \leftarrow \emptyset; M = S e l f A t t n (E)$
3
$f o r k = 1 \dots K_{m a x} a n d n o t c o n v e r g e d (C^{*}) :$
4
$P = S o f t m a x (D e c o d e r (E, V_{u})); X_{n e w} = T o p N (P ⊙ M)$
5
$V_{u} + = {x \in X_{n e w} | C r o s s A t t n (x, V_{g}) \geq τ}$
6
$w h i l e Q_{v} \neq \emptyset :$
7
$v = p o p_{m i n} (Q_{v}, f (v) \cdot σ (α)); N = T r a n s f o r m e r_{E x p a n d (v, E)}$
8
$Q_{E} + = {(v, w) | w \in N}; f o r e a c h (i, j) \in Q_{E}$
9
$ĉ = R e g r e s s o r (h_{i}, h_{j}); i f g (i) + ĉ < g (j) : p a r e n t (j) = i; u p d a t e h_{j}; u p d a t e C^{*}, π^{*} i f j \in V_{g}$
10
$θ \leftarrow θ - η \nabla {‖ĉ - c_{g t}‖}^{2}; r e s e t Q_{E}; r e o r d e r (Q_{v})$
11
$r e t u r n C^{*} \leq C_{A} ? P a t h S m o o t h (π^{*}, S e l f A t t n) : Ø$