Trajectory Planning and Control of Serially Linked Robotic Arm for Fruit Picking Using Reinforcement Learning

Fruit picking is a process in which a serial-linked robotic arm uses an end-effector for grasping fruit to minimize human effort, workload, accuracy, and efficiency. The real-world environment is quite unstructured and complex. It is a difficult task to detect the target fruit due to the presence of branches everywhere. To ensure the efficient navigation of a robot towards a detected fruit, creating a continuous action space is essential. This approach is more effective than relying on discrete action spaces. However, this is a challenging task. The real-world environment is too intricate to be accurately modelled mathematically due to its inherent complexity. All the prior methods used are either not intelligent and learning-based, or if they are intelligent, they are only useful for discrete action spaces. To achieve the required goal, Artificial Intelligence (AI) has been used in the given work to design a learning-based controller that will enable the robot to effectively reach its target fruit in continuous action spaces. For that, a prototype of the mechanical arm is trained using the Actor-Critic method named Proximal Policy Optimization (PPO), which is an on-policy and Model-Free Reinforcement Learning method. It is an optimization method; its main objective is to find optimal sets of inputs that can give us the maximum outputs. The results showed the success of the proposed method with an accuracy of 88.43%.