This article proposes a Reinforcement Learning (RL) driven autonomous Active Debris Removal (ADR) strategy based on Angles-Only Navigation (AON) theory. The policy network trained by the RL algorithm receives angles measurements as input, and outputs impulse information, thus combined the orbit determination solely based on angles measurements and autonomous approaching. To solve the low observability problem of AON, impulse maneuver is considered to gain observability. Thus, the network processes the coupled relationship of angles measurements and approaching maneuver, which balances the need to approach target and gain AON observability. This control method based on RL has the advantage of high autonomy and fast computation. Besides, the initial state of training is randomly decided, which effectively enhanced the generalization capabilities of the trained model, so the controller can handle more variable situations. Finally, simulation is performed to examine the performance of the policy network controller. The results showed great effectiveness and robustness in various environments.