One-Shot Learning of Multi-Step Tasks from Observation via Activity Localization in Auxiliary Video (ICRA 2019)

Wonjoon Goo and Scott Niekum, University of Texas at Austin

This repository contains codes for the ICRA 2019 paper. If you use this code as part of any published research, please consider referring the following paper.

@inproceedings{Goo2019,
  title     = {"One-Shot Learning of Multi-Step Tasks from Observation via Activity Localization in Auxiliary Video"},
  author    = {Wonjoon Goo and Scott Niekum},
  year      = {2019},
  booktitle = {2019 IEEE International Conference on Robotics and Automation (ICRA)},
  tppubtype = {inproceedings}
}

Setup Env

The code and following instructions are tested on Ubuntu 16.04.

Download Code

$ git clone https://github.com/hiwonjoon/ICRA2019-Activity-Localize.git --recursive

Prepare Conda Environment

$ conda env create -f environment.yml
$ conda activate icra2019_actlocal
$ pip install -r requirements.txt

Prepare Roboschool

$ cd roboschool
$ export ROBOSCHOOL_PATH = $(pwd)
$ mkdir bullet3/build
$ cd bullet3/build
$ cmake -DBUILD_SHARED_LIBS=ON -DUSE_DOUBLE_PRECISION=1 -DCMAKE_INSTALL_PREFIX:PATH=$ROBOSCHOOL_PATH/roboschool/cpp-household/bullet_local_install -DBUILD_CPU_DEMOS=OFF -DBUILD_BULLET2_DEMOS=OFF -DBUILD_EXTRAS=OFF  -DBUILD_UNIT_TESTS=OFF -DBUILD_CLSOCKET=OFF -DBUILD_ENET=OFF -DBUILD_OPENGL3_DEMOS=OFF ..
$ make -j4
$ make install
$ cd ../..
$ pip install -e $ROBOSCHOOL_PATH

Train

1. Prepare Datset

$ export DISPLAY:=0
$ python reacher_gen.py

2. Train Classifier-based Activity Localizer

$ python reacher_classify.py --LOG_DIR ./log/reacher/classify --DATA_TRAJS ./datasets/multi-reacher-easy/given.pkl

3. Train MAML-based Activity Localizer

$ python reacher.py --LOG_DIR ./log/reacher/maml --DATA_TRAJS ./datasets/multi-reacher-easy/given.pkl --META_TRAJS ./datasets/multi-reacher-easy/meta.pkl --PRETRAINED_MODEL ./log/reacher/classify/model.ckpt-2000

4. Train Reward Function

$ export DISPLAY:=0
$ python reacher_gen.py --path ./datasets/rwd_learn/multi-reacher-easy --num_trajs_per_task 1000
$ python train_r_func.py --LOG_DIR ./log/reward/perfect --DATA_TRAJ ./datasets/rwd_learn/multi-reacher-easy/given.pkl # Assume perfect localization
$ python train_r_func.py --LOG_DIR ./log/reward/maml --DATA_TRAJ ./datasets/rwd_learn/multi-reacher-easy/given.pkl --ALIGNER 'maml' #maml localization

5. Train a policy

$ export DISPLAY:=0
$ python train_policy.py --log_dir ./log/policy/maml --reward_type inferred --reward_model ./log/reward/maml/last.ckpt