torcs-node-client

torcs-node-client is Node.js client for running autonomous driver bots on The Open Race Car Simulator (TORCS).

Click image for Youtube Video

The Basics

TORCS is designed to enable pre-programmed AI drivers to race against one another, while allowing the user to control a vehicle using either a keyboard, mouse, or wheel input. In 2013, an enhancement was created to extend TORCS to support externally controlled robot drivers as part of the Simulated Car Racing (SCR) Championship. In this client-server configuration bots connect as clients to a TORCS simulator server.

TORCS with the SCR patch runs standalone in its own process to which this client connects and communicates over UDP on default port 3001. Every 20 ms TORCS sends a packet of sensor data for the simulated car and waits for commands. If no command is provided in the timeframe, the previous command is performed. This loop repeats until the race completes or your car stalls for > 5 seconds or your car crashes and sustains damage greater than the maximum damage setting.

The torcs-node-client enables a bot implemented in JavaScript (TypeScript) to connect to a TORCS server using the SCR network protocol and control a race car. While the default driver provided by this client is based on the Pure Pursuit algorithm it is very easy to implement and plugin your own driver controller, see Driver.ts for the api.

Prerequisites

You will need the following software installed:

• TORCS 1.3.7 or greater with the SCR patch installed.

• GIT commandline tools or the equivalent

• Node.js version 12 or greater

• If you run TORCS on a separate machine you will need the machine's ip. On Windows machines use ipconfig to access it's IP address. On Linux use ip a for it's IP address.

Getting Started

We will be running two process, the TORCS simulator and our driver bot. Begin by opening a command shell and cd to the folder that will contain your version of the torcs-node-client.

Clone this git repo

git clonde https://github.com/wayneparrott/torcs-node-client.git

Now let's build the package using the following commands:

cd torcs-node-client
npm i
npm build

Now launch the torcs-node-client.

node dist/index.js

or

node dist/index.js --host <ipaddr> --port <default=3001>

Open a command shell and start TORCS.

torcs

From the TORCS graphical UI configure a race and your car as follows:

Race > Quick Race > Configure Race > Accept > scr_server 1 driver > Accept > New Race

TORCS will remember the race, driver and car settings for future races and you should start the race using this sequence:

Race > Quick Race > New Race

At this point, the torcs-node-client will connect to TORCS and begin sending steering, accelation, brake and gear shift commands.

Restart

By default the torcs-node-client will automatically restart a race when it's race car suffers excessive damage. Additionally you can manually restart a race by entering a r key in the command shell.

Pure Pursuit Driver

The PurePursuitDriver.ts is a basic robot driver that uses a variable lookahead distance that is a function of the car's velocity and the longest sensor range reading. The cross-track error angle is angle of the long sensor range reading. On long straights an drift component detects if the car is too close the track edge and will slight nudge it back over with an additional steering input.

The velocity and acceleration control (accelerator, brake and gear) are computed together using. A PD controller is computes an acceleration between [-1,1]. This is then converted to accelerator and brake inputs. Both instantaneous maximum accelerator and brake inputs are limited to minimize wheel spin and skidding. Braking is controlled when the car enters a braking zone. A braking zone is resides in front of each turn usually beginning around 75 meters and extending into the turn. Within a braking zone the the target speed is set to the minimun of longest sensor range reading or a minimun car speed, e.g., 50 kph.

The car's speed is continuously monitored in order to detect when it is stalled or blocked in such a way that prevents it from moving for more than 5 seconds. After 5 seconds of no movement, the car is assumed to be permanently non-racable and the race is restarted.

Areas for Improvement

The result is a reasonably performant algorithm with additional room for such a simple controller. Because the steering controller uses the longest sensor range reading the car has a tendency to drive directly towards the apex of a turn rather than curve in to it. Additionally, over braking occurs on some turn segments. In such cases coasting or more liberal acceleration control would improve corner speed. A big future enhancement is to introduce a secondary controller that monitors and automatically tunes additional steering and acceleration parameters as the car learns the track.

Resources

1. TORCS
2. SCR Manual
3. SCR code repo
4. TORCS with SCR Patch
5. Pure Pursuit Algorithm
6. SCR sensor input and control output cheatsheet

Credits

Thanks to the CodeMix team at Genuitc. This software was developed using the CodeMix plugin for Eclipse.