One of the major problems in automatic target recognition (ATR) involves the recognition of images in different orientations. In a classical training-testing setup for an ATR for rotated targets using neural networks, the recognition is inherently static, and the performance is largely dependent on the range of orientation angles in the training process. To alleviate this problem, we propose a reinforcement learning (RL) approach for the ATR of rotated images. The RL is implemented in an adaptive critic design (ACD) framework wherein the ACD is mainly composed of the neuro-dynamic programming of an action network and a critic network. The proposed RL provides an adaptive learning and object recognition ability without a priori training. Numerical simulations demonstrate that the proposed ACD-based ATR system can effectively recognize rotated target with the whole range of 180 • rotation. Analytic characterization of the learning algorithm provides a sufficient condition for its asymptotic convergence. Finally, we obtain an upper bound for the estimation error of the cost-to-go function under the asymptotic convergence condition.