Collecting Human Demonstrations
We provide teleoperation utilities that allow users to control the robots with input devices, such as the keyboard and the SpaceMouse. Such functionality allows us to collect a dataset of human demonstrations for learning. We provide an example script to illustrate how to collect demonstrations. Our collect_human_demonstrations script takes the following arguments:
directory:path to a folder for where to store the pickle file of collected demonstrations
environment:name of the environment you would like to collect the demonstrations for
device:either “keyboard” or “spacemouse”
Note that the rendering window must be active for these commands to work.
|spacebar||toggle gripper (open/close)|
|w-a-s-d||move arm horizontally in x-y plane|
|r-f||move arm vertically|
|z-x||rotate arm about x-axis|
|t-g||rotate arm about y-axis|
|c-v||rotate arm about z-axis|
3Dconnexion SpaceMouse controls
|Right button||reset simulation|
|Left button (hold)||close gripper|
|Move mouse laterally||move arm horizontally in x-y plane|
|Move mouse vertically||move arm vertically|
|Twist mouse about an axis||rotate arm about a corresponding axis|
Replaying Human Demonstrations
We have included an example script that illustrates how demonstrations can be loaded and played back. Our playback_demonstrations_from_hdf5 script selects demonstration episodes at random from a demonstration pickle file and replays them.
We have included some sample demonstrations for each task at
Our sister project RoboTurk has also collected several human demonstration datasets across different tasks and humans, including pilot datasets of more than a thousand demonstrations for two tasks in our suite via crowdsourcing. You can find detailed information about the RoboTurk datasets here.
Structure of collected demonstrations
Every set of demonstrations is collected as a
demo.hdf5 file. The
demo.hdf5 file is structured as follows.
date (attribute) - date of collection
time (attribute) - time of collection
repository_version (attribute) - repository version used during collection
env (attribute) - environment name on which demos were collected
demo1 (group) - group for the first demonstration (every demonstration has a group)
model_file (attribute) - the xml string corresponding to the MJCF mujoco model
states (dataset) - flattened mujoco states, ordered by time
actions (dataset) - environment actions, ordered by time
demo2 (group) - group for the second demonstration
(and so on)
The reason for storing mujoco states instead of raw observations is to make it easy to retrieve different kinds of observations in a postprocessing step. This also saves disk space (image datasets are much larger).
Using Demonstrations for Learning
We have recently released the robomimic framework, which makes it easy to train policies using your own datasets collected with robosuite, and other publically released datasets (such as those collected with RoboTurk). The framework also contains many useful examples for how to integrate hdf5 datasets into your own learning pipeline.
The robosuite repository also has some utilities for using the demonstrations to alter the start state distribution of training episodes for learning RL policies - this have proved effective in several prior works. For example, we provide a generic utility for setting various types of learning curriculums which dictate how to sample from demonstration episodes when doing an environment reset. For more information see the
We have verified that deterministic action playback works specifically when playing back demonstrations on the same machine that the demonstrations were originally collected upon. However, this means that deterministic action playback is NOT guaranteed (in fact, very unlikely) to work across platforms or even across different machines using the same OS.
While action playback trajectories are quite similar even if not completely identical to the original collected state trajectories, they do tend to drift over time, and should not be relied upon to accurately replicate demonstrations. Instead, we recommend directly setting states to reproduce the collected trajectories, as shown in playback_demonstrations_from_hdf5.