Working with MimicGen in MimicLabs#

We provide a detailed description of how we automate integration with MimicGen, and how you can tune it for your work.

Some important scripts are:

Preparing source demonstrations: mimiclabs/mimicgen/scripts/prepare_src_dataset.py
Creating datagen configs: mimiclabs/mimicgen/scripts/generate_configs_and_jobs.py
Generating datasets: mimiclabs/mimicgen/scripts/generate_dataset.py

Preparing your source demonstrations#

The prepare_src_dataset.py script is used to add object-centric information to source demonstrations before they can be used for dataset expansion. Below are all paramaeters that can be changed in the this script.

Argument	Description
dataset	path to source hdf5 dataset (after post-processing), which will be modified in-place
env_interface	interface class to use for this source dataset. default is `MG_MimicLabs`. We built this class to extract object-centric info about demos using `demonstration` states in BDDL
env_interface_type	type of environment interface; we set it to `robosuite` by default and is provided by MimicGen
n	only process that many trajectories (used for debugging)
filter_key	filters the source demos to be `source_hdf5_file[f"mask/{filter_key}"]`
output	path to output hdf5; if not specified, dataset is modified in place

The MG_MimicLabs interface class#

Preparing source demonstrations relies on a MimicGen environment interface that we define for all tasks described in MimicLabs, called MG_MimicLabs. It is a subclass of the RobosuiteInterface class in MimicGen, and overrides its get_object_poses() and get_subtask_term_signals() functions. We define this interface class in mimiclabs/mimicgen/env_interface.py.

This class uses the demonstration states in the BDDL to convert them into subtask keys that are added to source_hdf5_file[f"data/demo_i/datagen_info"]. For example, for the given demonstration states,

(:demonstration
    (:Open wooden_cabinet_1_top_region)
    (:Grasp object_1)
    (:In object_1 wooden_cabinet_1_top_region)
)

the following key-value pairs are added for each timestep

"subtask_1_open_wooden_cabinet_1_top_region": eval("open", "wooden_cabinet_1_top_region")
"subtask_2_grasp_object_1": eval("grasp", "object_1")
"subtask_3_object_1_wooden_cabinet_1_top_region": eval("in", "object_1", "wooden_cabinet_1_top_region")

where eval() is a function that evaluates any predicate in the scene. Implementations for different predicate can be found in mimiclabs/mimiclabs/envs/predicates/predicates.py which imports multiple implementations from LIBERO.

Creating MimicGen configs and generation jobs#

The generate_configs_and_jobs.py is used to create configs for data generation with MimicGen.

This script calls the generate_mg_config_classes() function in mimiclabs/mimicgen/config.py to generate config classes. The name of each class is the BDDL filename, with a TYPE attribute assigned to be the task suite name. This function dynamically creates config classes that inherit from MG_Config and registers them with the ConfigMeta metaclass. The task_config() and obs_config() methods are populated based on the parsed BDDL files and args to this function respectively, allowing for flexible task and observation configurations.

Below are some useful arguments to this script:

Argument	Description
task_suite_name	name of task suite to generate demos for
source_demos_dir	directory containing source demos for the task suite, one subdirectory per task bddl file
generation_dir	directory to store generated demos for the task suite, one subdirectory per task bddl file
num_demos	number of demos to generate for each task bddl
camera_names	list of camera names to use for rendering and store to dataset; default is [“agentview”, “robot0_eye_in_hand”]
camera_height	camera height to use for rendering; default is 84
camera_width	camera width to use for rendering; default is 84

Generating datasets#

The generate_dataset.py is finally used to start data generation using MimicGen given a generated config. An important thing to note is that we call the generate_mg_config_classes() from mimiclabs/mimicgen/config.py in this script to create and register config classes for each BDDL in the task suite.