Large-scale Robot Learning Dataset of RoboNet

2019-12-09

In the past decade, we have seen deep learning systems provide transformative solutions to various perception and reasoning problems, such as object recognition in images and speech recognition and translation.


While deep reinforcement learning methods can learn multiple skills, training them with large amounts of data from the target environment is challenging. Applying data-driven approaches to robotics not only requires the development of powerful reinforcement learning methods, but also access to large and diverse robot datasets.


Unfortunately, it has been proven challenging to design and adopt large datasets in reinforcement learning and robotics. Each robotics lab has its own hardware and experimental setup, making it difficult to provide useful robot technology datasets to each lab. Therefore, we created RoboNet, a scalable and diverse robot interaction dataset that collects data from four different research labs. The collaborative nature of this work allows us to easily capture a variety of data across various lab settings, objects, robot hardware, and camera viewpoints.



Collecting RoboNet


RoboNet consists of 15 million video frames collected by different robots interacting with various objects in a tabletop environment. Each frame includes images recorded by the robot's camera, arm poses, force sensor readings, and gripper states. The collection environment, including camera views, the appearance of the table or trash can, and objects in front of the robot, varies between each experiment. As the collection is fully autonomous, a large amount of data can be collected inexpensively across multiple institutions.



How do we use RoboNet?


After collecting a diverse dataset, we will investigate how to use it to transfer general skill learning to new environments through experimental research. First, we pretrain visual dynamics models on subsets of data from RoboNet, and then fine-tune them to work in unseen test environments with a small amount of new data. The constructed test environments include different lab setups, new cameras and viewpoints, extended robots, and novel objects purchased after data collection.


After fine-tuning, we deploy the learned dynamics models into the test environments using reinforcement learning algorithms based on visual foresight to perform control tasks such as object grasping and placing. Now, we can quantitatively evaluate whether our pretrained controllers can learn skills in new environments faster than randomly initialized controllers. In each environment, we use a set of standard benchmark tasks to compare the performance of the pretrained controllers with models trained only on data from the new environment. The results show that the fine-tuned models have about 4 times higher likelihood of completing benchmark tasks compared to models trained without RoboNet. Furthermore, the pretrained models even slightly outperform models trained on a large amount of data obtained solely from the test environment (5-20 times). This indicates that transfer from RoboNet indeed provides significant performance improvement compared to training from scratch!


This work takes the first step towards creating learning robotic agents that can operate in a wide range of environments and different hardware. While our experiments primarily explore model-based reinforcement learning, we hope that RoboNet will inspire a broader robotics and reinforcement learning community to research how to extend model-based or model-free RL algorithms to meet the complexity and diversity of the real world.

RECENT AI NEWS

RECENT AI TOOLS