Agents

A set of pre-trained algorithmic baseline agents across various versions and environment suites are available with each release of RoboHive.

RoboHive has out-of-the-box support for multiple hardware including teleoperation devices. Leveraging this support, we have engaged human agents to collect RoboSet -- RoboHive's dataset module. It's a collection of datasets (both algorithmic and well as human teleOp) collected with compatibility with the RoboHive ecosystem and different environments suites. RoboSets, especially human-teleOp data, are of high quality and can be used to recover imitation learning or offline learning-based baseline agents. (As an example, we present a set of pre-trained visual imitation learning agents below. )

RoboHive can be used with any openAI-Gym compatible algorithmic baselines to train agents for its environment. To bootstrap the ecosystem, RoboHive also maintains AgentHive -- a collection of tools and helper scripts for training agents as well as offline and online execution of pre-packaged trained agents.

To demonstrate the comprehensiveness of the RoboHive ecosystem, we bring together 

• state-of-the-art visual encoder support from RoboHive.

• pre-packaged dataset included in RoboSet.

• baselines algorithms from AgentHive. 

to present a head-to-head comparison of various state-of-the-art visual representation learning methods. We provide an overview of the results below. Please visit the Visual IL baseline link on the left for details and reproducibility.