RobotSlang Benchmark README

Use the following script to setup a virtual environment and cache shortest paths and particle measurements.

source install.sh 

Training

Use the provided train.py script to train the baseline seq2seq model. The script takes the following options:

Option	Possible values
feedback	'sample', 'teacher'
eval_type	'val', 'test'
blind	'vision', 'language', ''

Note that in the traditional experiment where no blinding occurs, the blind parameter is blank. As an example, to train a model with teacher forcing, use the following command.

python train.py --feedback=teacher --eval_type=val --lr=0.0001

Note that your device must be CUDA-enabled.

Visualizations

To make videos of our trials using the simulator run the following:

python make_simulation_videos.py 

Raw Data

Download videos of robotslang trials here

Questions?

Contact us at shurjo@umich.edu