mynkpl1998 / upgraded-octo-lamp

An autonomous driving simulator for modelling Vehicle to Infrastructure (V2I) conditions.

V2I

V2I is an open-source simulator for autonomous driving research developed at Cyber-Physical Systems Lab IIIT-Delhi. It has been developed to address the need of a simulator which can simulate Vehicle-To-Infrastructure communication. The simulator is designed primarily to learn driving and querying policies using Reinforcement Learning.

Features

The simulator is under heavy development, new features are being added everyday. For now, the checked box represents the implemented features and rest will be implemented soon.

• Occupancy Grids
• Velocity Grids
• Multi Lane
• Driving Model for Non-Autonomous Vehicles - Intelligent Driver Model
• Multi Agent

Installation

We recommend you to use virtual environment like conda for installation. You can get the mini version of conda from here. We will use this MiniConda as our base python installation and will do the required installation on top of that. To install the dependencies, run the following commands inside the conda environment

conda create -n v2i python=3.6.8
# Activate the environment
conda activate v2i

# Install all required dependencies
pip install -r requirements.txt

# Install Tensorflow
conda install tensorflow-gpu

upgraded-octo-lamp's Issues

Add frame skip module to the simulation

[Doc] document gen_trajectories, gen_tfTrajectories.py

simulator reaches terminal states if though no collision takes place.

Configuration to reproduce the result.

Script : test.py

Sim Config

config:
  # Simulation Parameters
  t-period: 0.1 # in secs
  
 # IDM settings
   max-speed: 10 # in m/s
   local-view: 20 # in metres (one side)
   cextended-view: 20 # in metres (one side)
   reg-size: 5 # in metres
   cell-size: 1 # in metres

  # Reward Function
  collision-penalty: -10

  # UI settings
  render: true
  fps: 5

  # Set Mode
  mode: 'test'

  # Frame Skip
  frame-skip-value: 1

Occupancy Grid detection is not working properly near ego vehicle in the lane next to ego vehicle lane

Add support for APE-X

need to manually set "mode" to train or test to enable/disable rendering when using train.py or rollouts.py. This should be done automatically

train and roll-outs script requires absolute path to the simulation configuration file

Possibility of using weighted sampling for during training episode density.

It has been observed that, for densities > 0.4, significant number of lane changes leads to collision, which might be a reason for agent not to select lane change action. May be we can try sampling low density episodes more often high density to mitigate this problem.

Add support to control planning action using keyboard.

This will allow to easily debug the simulator.

Add support to generate performance data with enabling/disabling tf Lights [rollouts.py]

invalid density value error on setting density manually to 0.3 or 0.7.

Reward Shaping

With speed based reward function, average length of the episode is low for higher density because the cars gets too close to each other and using a stochastic policy leads to selection of action which results in collision.

Solution : Penalise the agent if bumper to bumper distance becomes smaller than a threshold distance. (like 2m)

add detailed step by step examples of simulator

Examples to include

1. Local View
2. With Comm
3. With Comm with different size of regions

Add support to change lane density via configFiles

Add support to load model from checkpoint and control the agent via external input [play.py]

• If user doesn't provide any control input then agent action comes from trained policy else policy action is overwritten by user input.

• Add option to enable disable traffic light and manually controlling traffic lights from external input.

Implement frame stacking to emulate k-order MDP

Add support for trafficDensity=0.0.

[rollouts.py] bar plot of probability of actions is not updating correctly.

add red light to simulate mix traffic.

As discussed in the meeting, the ego vehicle should not travel faster than the what is allowed by the local view size. However this is not current the case as we are not simulating the mixed traffic correctly. Therefore, Add a light at a the junction for the same.

Add lane change support

• Implement MOBIL model
• https://mtreiber.de/publications/MOBIL_TRB.pdf

current lane change behaviour results in wrong policies.

According to lane change behaviour which is

1. If lane change is valid move to next lane, but no change is position and speed.

This causes wrong policies like when tf lights are enabled and both lanes are blocked agent uses lane change action to prevent agent to move forward.

Revert back to correct lane change behaviour which is in addition to 1 includes doing a do-nothing action.

Add Reward Function, Generalise this using yaml functions

add training support using rllib

Add CI/CD for builds

Add Communication penalty in rewards.

Add a function which penalises for querying a region.
Create an independent function.

main() -> reward func() -> commPenalty() -> main()