End-to-End Deep Reinforcement Learning for Inclined Ladder Steps Grasping in Humanoid Robots

When it comes to grasping inclined ladder, the conventional approach involves using pre-determined fixed actions to control the humanoid robot. However, this method necessitates manual design and imposes strict initial position requirements on the robot. To overcome this challenge, we propose an autonomous grasping method for an humanoid robot using a deep reinforcement learning algorithm called Deep Q-Network (DQN). Our approach involves developing an end-to-end network model that takes camera images and servo angles as inputs and generates optimal action policies for the humanoid robot. By utilizing this strategy, the humanoid robot achieves a high success rate in grasping the inclined ladder. To verify the effectiveness of our method, we conducted performance tests on the model in various scenarios and compared it with the fixed action control method.