Visual Reaction: Learning to Play Catch with Your Drone

By Kuo-Hao Zeng, Roozbeh Mottaghi, Luca Weihs, and Ali Farhadi

Paper | Video | BibTex

We address the problem of Visual Reaction, where the idea is to forecast the future and plan accordingly. We study the task in the context of catching objects with a drone. An object is thrown in the air, and the drone should plan to catch it. Each object has different physical properties and might collide with other objects and structures in the scene, making the task quite challenging.

Citing

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

@inproceedings{khz2020visualreaction,
  author = {Zeng, Kuo-Hao and Mottaghi, Roozbeh and Weihs, Luca and Farhadi, Ali},
  title = {Visual Reaction: Learning to Play Catch with Your Drone},
  booktitle = {CVPR},	    
  year = {2020}
}

Set Up

0. Requirements

   We implement this codebase on Ubuntu 18.04.3 LTS and also have tried it on Ubuntu 16.

   In addition, this codebase needs to be executed on GPU(s).

1. Clone this repository

   git clone git@github.com:KuoHaoZeng/Visual_Reaction.git
   

2. Intsall xorg if the machine does not have it

   Note: This codebase should be executed on GPU. Thus, we need xserver for GPU redering.

   # Need sudo permission to install xserver
   sudo apt-get install xorg
   

   Then, do the xserver refiguration for GPU

   sudo python startx.py
   

3. Using python 3.6, create a venv

   Note: The python version needs to be above 3.6, since python 2.x may have issues with some required packages.

   # Create venv and execute it
   python -m venv venv && source venv/bin/activate
   

4. Install the requirements with

   # Make sure you execute this under (venv) environment
   pip install -r requirements.txt
   

Environment/Dataset

We extend AI2-THOR by adding a drone agent and a luncher to play catch. After the luncher throws an object, the drone needs to predict the trajectory of the object from ego-centric observations and move to a position that can catch the object.

We collect a dataset consiting of 30 scenes (Living Rooms) in AI2-THOR. The initial positions of the drone and the luncher are random. The luncher randomly select an object from 20 objects list, and throws it with random magnitudes in random directions. Overall, we collect 20K training trajectories, 5K validation trajectories, and 5K testing trajectories.

The training data, validation data, and testing data are available in the data folder.

For more information about how to control the drone agent in the environment, please vist this Doc.

Generate your own data

TBD

After the data generation, you need to change the data_dir to your data folder in the config file:

...
base_dir: "results/{{exp_prefix}}"
data_dir: "data" <-- change it to your data folder.
...

Play Catch with Your Drone!

Note: You can always change or adjust the hyperparameters defined in the config file to change the setting such as how many GPUs are going to be used, how many threads are going to be used, how often you want to store a checkpoint, etc. You can also change the learning rate, number of iterations, batch size, etc. in the config file.

Test the pretrained model on validation/testing set

# Download the pretrained model
wget https://homes.cs.washington.edu/~khzeng/Visual_Reaction/pretrained.zip
unzip pretrained.zip && mv pretrained/* results/ && rm pretrained.zip

# Test the model
# The testing results would be stored by jsonlines file
# For the forecaster only
python main.py --config configs/pretrained_forecaster_test.yaml
# For the forecaster + action_sampler
python main.py --config configs/pretrained_action_sampler_test.yaml

# Evaluate the results
# For the forecaster only
python eval.py results/pretrained_forecaster test
# For the forecaster + action_sampler
python eval.py results/pretrained_action_sampler test

Train a new forecaster

# Train
python main.py --config configs/forecaster_train.yaml

# Validate or Test
python main.py --config configs/forecaster_val.yaml
python main.py --config configs/forecaster_test.yaml

# Eval
python eval.py results/forecaster val
python eval.py results/forecaster test

Train a new action sampler with a trained forecaster

# Prepare the trained forecaster
cd results
mkdir action_sampler
mkdir action_sampler/checkpoints
cp -r forecaster/checkpoints/$FORECASTER_YOU_LIKE_TO_USE action_sampler/checkpoints/0000000
cd ..

# Train
python main.py --config configs/action_sampler_train.yaml

# Validate or Test
python main.py --config configs/action_sampler_val.yaml
python main.py --config configs/action_sampler_test.yaml

# Eval
python eval.py results/action_sampler val
python eval.py results/action_sampler test

Main Results

Model	Success Rate
Forecastor w/ uniform AS (ours)	26.0 ± 1.3
Forecastor w/ action sampler (ours)	29.3 ± 0.9
CPP w/ KF	23.2 ± 1.3
CPP	22.9 ± 2.3

Different Mobility

Noisy Movement