A tutorial for DSRNN-Crowd Navigation

This is a tutorial for installing DSRNN-Crowd Navigation.
NOTE : Download the modules from URL, not this repository.

For a detailed explanation

1. How to install

Install Python 3.6

The code may work with other versions of Python, but Python 3.6 is highly recommended.

Make sure that you need to use Python 3.6 as default, so you might have to use Virtualenv or Anaconda.
Clone the repository and install the required python package using pip.
```
pip install -r requirements.txt
```

Install OpenAI Baselines.

git clone https://github.com/openai/baselines.git
cd baselines
pip install -e .

Install Python-RVO2 library.

git clone https://github.com/sybrenstuvel/Python-RVO2
cd Python-RVO2
pip install -r requirements.txt
python setup.py build
python setup.py install

Finally, you have to check if you can import installed modules.
If there's no error messages, modules are installed correctly.
```
python
>> import tensorflow
>> import pandas
>> import baselines
>> import rvo2
```

If all modules are imported correctly, the next step is training & testing DS-RNN.

2. How to use

Getting started

This repository is organized in three parts:

crowd_sim/ folder contains the simulation environment.
crowd_nav/ folder contains configurations and non-neural network policies.
pytorchBaselines/ contains the code for the DSRNN network and ppo algorithm.
Below are the instructions for training and testing policies.

Change configurations

Environment configurations and training hyperparameters: modify crowd_nav/configs/config.py
For FoV environment (left in the figure below): change the value of robot.FOV in config.py
For Group environment (right in the figure below): set sim.group_human to True in config.py

Run the code

Train a policy.
```
python train.py
```
Test policies.

Please modify the test arguments in the beginning of test.py.

Two trained example weights for each type of robot kinematics are provided:
- Holonomic: data/example_model/checkpoints/27776.pt
- Unicycle: data/example_model_unicycle/checkpoints/55554.pt
```
python test.py
```
Plot training curve.
```
python plot.py
```

3. Results

Execute .py files

Trained about 10,000,000 timesteps for holonomic case.

Updates 0, num timesteps 360, FPS 637 
 Last 2 training episodes: mean/median reward -16.7/-16.7, min/max reward -19.0/-14.5

Updates 20, num timesteps 7560, FPS 645 
 Last 83 training episodes: mean/median reward -13.4/-15.9, min/max reward -36.1/15.4
 
 ...

Testing policies

2022-06-14 10:11:01, INFO: TEST  has success rate: 0.90, collision rate: 0.10, timeout rate: 0.00, nav time: 10.06, total reward: 20.8718
2022-06-14 10:11:01, INFO: Frequency of being in danger: 0.02 and average min separate distance in danger: 0.15
2022-06-14 10:11:01, INFO: TEST  has average path length: 190.30, CHC: 11.43
2022-06-14 10:11:01, INFO: Collision cases: 11 15 19 26 36 51 72 80 93 108 118 123 132 137 154 159 162 173 181 187 193 198 223 225 231 257 267 285 286 287 289 297 300 319 322 323 355 357 359 361 368 375 390 404 424 442 445 460 464 468
2022-06-14 10:11:01, INFO: Timeout cases: 9
 Evaluation using 500 episodes: mean reward 22.35157

Plotting training curve

Code Review

Before training

In crowd_nav/configs/config.py, you can change parameters for training.

environment settings
human settings : size of human, maximum velocity, FoV, policy to control human
robot settings : size of robot, maximum veloity, FoV
reward function : rewards for your model
training config : path for saving new model file or loading created model file, number of timesteps

After training

In data/YOUR_DATA/progress.csv and data/YOUR_DATA/checkpoints/YOUR_MODEL.pt, you can find trained data.

eprewmean : mean of reward for 1 update
policy_entropy : randomness of actions an agent can take
policy_loss : The mean magnitude of policy loss function. Correlates to how much the policy (process for deciding actions) is changing. The magnitude of this should decrease during a successful training session. These values will oscillate during training. Generally they should be less than 1.0.
value_loss : The mean loss of the value function update. Correlates to how well the model is able to predict the value of each state. This should increase while the agent is learning, and then decrease once the reward stabilizes. These values will increase as the reward increases, and then should decrease once reward becomes stable.

jhforstudy / DSRNN_CrowdNav