Ask4Help: Learning to Leverage an Expert for Embodied Tasks

By Kunal Pratap Singh, Luca Weihs, Alvaro Herrasti, Jonghyun Choi, Aniruddha Kembhavi, Roozbeh Mottaghi

Paper | BibTex | Talk

This is the official code base for our paper presented in NeurIPS, 2022.

Installations

1. Clone this repository

   git clone https://github.com/allenai/ask4help.git

2. Install the requirements in virtual environment using the following commands

   pip install -r requirements.txt
   pip install -r dev_requirements.txt
   pip install -r allenact_plugins/ithor_plugin/extra_requirements.txt
   pip install -r allenact_plugins/robothor_plugin/extra_requirements.txt

3. You will require xorg on your machine to run AI2THOR in headless mode. You can install it using

sudo apt-get update
sudo apt-get install xorg

4. You can start xorg using.

ai2thor-xorg start

Training an Ask4Help Policy

We build upon the RoboTHOR object navigation training code in the Allenact framework.

To train an Ask4Help Policy run the following command

python main.py 
-b projects/objectnav_baselines/experiments/robothor/clip
objectnav_robothor_rgb_clipresnet50gru_finetune_ask_ddppo.py 
-o storage/ask4help_train 
--checkpoint ./storage/exp_Objectnav-RoboTHOR-RGB-ClipResNet50GRU-DDPPO_embedded_fix__stage_00__steps_000110213221.pt 
--extra_tag first_run 
--restart_pipeline

We also provide the config to training with multiple reward configurations.

python main.py 
-b projects/objectnav_baselines/experiments/robothor/clip
objectnav_robothor_rgb_clipresnet50gru_finetune_ask_adaptive_reward_ddppo.py 
-o storage/ask4help_train 
--checkpoint ./storage/exp_Objectnav-RoboTHOR-RGB-ClipResNet50GRU-DDPPO_embedded_fix__stage_00__steps_000110213221.pt 
--extra_tag adaptive_reward_run 
--restart_pipeline

Evaluation

We evaluate the ask4help policies in RoboTHOR validation environments. To run the evaluation:

python main.py
-b project/objectnav_baselines/experiments/robothor/clip
objectnav_robothor_rgb_clipresnet50gru_finetune_ask_ddppo.py 
-o storage/ask4help_train 
--checkpoint $PATH_TO_CKPT$ 
--extra_tag first_run_eval 
--eval

Qualitative Example

We present a qualitative example, where the agent is looking for laptop. Note how the agent struggles to get past the chair to get close enough to the laptop. The Ask4Help policy requests expert assistance, denoted by the progress bar turning red. The expert helps the agent get around the chair, and hence successfully end the episode.

FloorPlan_Val1_2_Laptop_4.mp4

Citation

If you find this project useful in your research, consider citing our work:

@inproceedings{singh2022ask4help,
  author = {Singh, Kunal Pratap and Weihs, Luca and Herrasti, Alvaro and Kembhavi, Aniruddha and Mottaghi, Roozbeh},
  title = {Ask4Help: Learning to Leverage an Expert for Embodied Tasks},
  booktitle = {NeurIPS},	    
  year = {2022}
}