Vision-Based Multi-Task Manipulation for Inexpensive Robots Using End-To-End Learning from Demonstration

In this paper, we propose a multi-task learning from demonstration method that works using raw images as input to autonomously accomplish a wide variety of tasks in the real world using a low-cost robotic arm. The controller is a single recurrent neural network that can generate robot arm trajectories to perform different manipulation tasks. In order to learn complex skills from relatively few demonstrations, we share parameters across different tasks. Our network also combines VAE-GAN-based reconstruction with autoregressive multimodal action prediction for improved data efficiency. Our results show that weight sharing and reconstruction substantially improve generalization and robustness, and that training on multiple tasks simultaneously greatly improves the success rate on all of the tasks. Our experiments, performed on a real-world low-cost Lynxmotion arm, illustrate a variety of picking and placing tasks, as well as non-prehensile manipulation.