Order2Taskplan-pytorch

This is a PyTorch implementation of the Episodic memory module described in our IROS 2018 paper "Adaptive Task Planner for Performing Home Service Tasks in Cooperation with a Human". Please note that this repository contains only the implementation of the episodic memory module(the deep learning part) in the paper.

Input

• A perceived object list:
  • An object in the list can be expressed in multiple words in natual language
• User command(order):
  • Natural language instruction or order that an user gives to a robot.(optional)

Output

• A robot behavior sequence(taskplan) with n behaviors:
  • subject₁ verb₁ object₁ subject₂ verb₂ object₂ ... subject_n verb_n object_n
  • subject_t = object_t if verb_t deals with only one object.

Dataset

• Total 50,000 input-output pairs (Train: 45,000, Test: 5,000)
• Ten robot behaviors: grasp, throw, locate, release, move, pour, put, push, sprinkle, and squeeze
• Every pair is related to one of these four scenarios:

• Example

Requirements

• Ubuntu 16.04+
• Python 3.6+
• Pytorch 0.3.1+
• Torchtext
• Numpy
• Matplotlib

Installing Order2Taskplan-pytorch

Order2Taskplan requires Python 3.6 or higher. It also requires installing PyTorch 0.3.1+ (warnings occur on 0.4.0). Its other dependencies are listed in requirements.txt. CUDA is required.

Run the following commands to clone the repository and install Order2Taskplan:

git clone https://github.com/chickenbestlover/Order2Taskplan-pytorch.git
cd Order2Taskplan-pytorch; pip install -r requirements.txt

Training

• Step 1: Train Order2Taskplan model:

CUDA_VISIBLE_DEVICES=0 python 1_train_order2taskplan_model.py

• Step 2: Train Hallucination model(optional, required if you don't want the user command(order) input):

CUDA_VISIBLE_DEVICES=0 python 2_train_hallucination_model.py 

• We also provide pretrained models: 1_best_model.pt, 2_best_model.pt

Download and place them to checkpoint/best_model

Demo Video

• This video was presented at IROS 2018.

Citation

Please cite the IROS 2018 paper if you use Order2Taskplan in your work:

@inproceedings{lee2018taskplan,
  title={Adaptive Task Planner for Performing Home Service Tasks in Cooperation with a Human},
  author={Lee, Seung-Jae and Park, Jin-Man and Kim, Deok-Hwa and Kim, Jong-Hwan},
  booktitle={Intelligent Robots and Systems (IROS), 2018 IEEE/RSJ International Conference on},
  year={2018}
  organization={IEEE}
}

License

The pre-trained models and the codes are released for uncommercial use.

Please open an issue if you have questions.