Many robotics applications, softwares, techniques and modules usually require optimizations of hyperparameters in order to be efficient for specific tasks, commonly performed offline by a human expert. In this work, we consider the use case of a grasping robot for industrial bin-picking. We propose a Developmental Cognitive Framework endowing the robot with the capability to self-explore and optimize efficiently by itself such parameters, even from noisy and expensive evaluations, during its own lifetime, after deployment. The robot cognitive architecture is based on reasoning mechanisms (in particular a Bayesian Optimization (BO) module in charge of the exploration) but also a Long-Term Memory (LTM, with episodic, semantic and procedural sub-memories) allowing the robot to take advantage of knowledge from similar past experiences in order to enhance the BO search (using Case-Based Reasoning paradigm applied with some transfer and meta-learning strategies). We evaluated the system with the constrained optimizations of 9 continuous hyperparameters for a professional software in simulated industrial robotic arm bin-picking tasks (a step that is needed each time to handle correctly new object) using only a very small optimization budget of 30 iterations. We show that the BO is significantly benefiting from the combination of meta (ML) and transfer (TL) learning strategies for each of the 10 objects tested to achieve very good performance in any case, despite a very little search budget (overall from 79.52%of success from vanilla BO to 83.96% with TL, 83.63% with ML and 85.62% with ML+TL).