Towards accelerated robotic deployment by supervised learning of latent space observer and policy from simulated experiments with expert policies

摘要

Up until today robotic tasks in highly variable environments remain very difficult to solve. We propose accelerated robotic deployment through task solving on low-level sensor data in simulation. A simulation allows for a lot of data, which is usually not available in a real world robotic setup due to cost and feasibility. Solving tasks in simulation is safe and a lot easier due to the huge amount of feedback from virtual sensory data. We present a novel sim2real architecture for converting simulated low level sensor data policies to high level real world policies. After solving a task we let the robot complete it a number of times in simulation using domain randomization, while doing so we save the simulated sensor data corresponding to the real robotic setup and actions taken. Given these sensor data and actions a task specific policy can be trained using our architecture. In this paper we work towards a proof of concept by simulating a simple low cost manipulator in pybullet to pick and place an object based on image observations.