Multi-Task Hierarchical Imitation Learning for Home Automation
This is code for the paper Multi-Task Hierarchical Imitation Learning for Home Automation, Roy Fox*, Ron Berenstein*, Ion Stoica, and Ken Goldberg, CASE 2019. It enables running the HIL-MT server, generating annotated demonstrations, learning hierarchical controllers, and rolloing learned controllers on an HSR robot.
In addition to the dependencies in requirements.txt, the
HSREnv depends on pyyolo.
Please follow these instructions to install pyyolo, and then change the paths in vision.py to your installation path.
To run the HIL-MT server:
To run the server in debug mode, set the
DEBUG flag to
True in server.py.
To rollout a demonstration of the
python rollout.py --domain dishes --task SetTable --data data --teacher
Available tasks are:
SetTable in the
dishes domain; and
Pyramid<n> (pyramid of height
n) in the
Annotated demonstrations will be saved in the path provided to
To implement new domains, inherit from
Env (or its subclass
HSREnv) -- see
PyramidEnv for example.
To implement new tasks, inherit from
Agent (or its subclass
HierarchicalAgent) -- see
PyramidAgent for example.
Human teleoperation is performed by actions starting with
Record_, which pause the script to allow human control and record it.
Before training, the recorded demonstrations must be fixed to format the recorded control as robot control.
To fix all demonstrations in a path:
python -c "import fix; fix.fix_record('data/dishes/SetTable_fix', 'data/dishes/SetTable_rec')"
Training hierarchical controller
To train controllers for the
python train.py --domain dishes --tasks ClearTable SetTable --data data
All controllers but the last will be trained independently from all their available data.
The last controller (
SetTable in this case) will be trained with detailed mode selection with the data of past tasks and with new demonstrations added one by one.
The results of training will be saved in
--runs <n> to repeat training for
n independent trials.
--independent to train the last controller only independently of the other tasks.
--full-batch to train the last controller from all available data, rather than stopping when enough demonstrations were given to successfully train and validate the controller.
Evaluating a trained controller
To rollout a trained controller:
python rollout.py --domain dishes --task SetTable --model model --data eval
Data from the evaluation experiment will be saved in the path provided to