Gym-Khana Documentation

Gym-Khana is a fast, deterministic Gymnasium environment for training autonomous 1/10th scale racing agents to race, drift, and recover. It supports RL training with SB3 and wandb, as well as classical controllers like MPC and Pure Pursuit.

GitHub repository: https://github.com/TeoIlie/Gym-Khana

Built on top of f1tenth_gym from UPenn (original docs).

Features

• Realistic vehicle dynamics: Kinematic, single-track, multi-body, and single-track drift (PAC2002 tire) models

• Fast simulation: Numba JIT-compiled dynamics for faster-than-real-time execution

• RL integration: SB3 and wandb support with parallel training via SubprocVecEnv

• Drift training: STD model with Pacejka tire physics for learning aggressive driving

• Recovery training: Train policies to stabilize an out-of-control vehicle with curriculum learning

• Deterministic: Seeded randomness and simultaneous agent stepping for reproducible experiments

• Multi-agent support: Multiple competing vehicles in the same environment

• LiDAR simulation: Accurate ray-casting for perception research

• Configurable observations & actions: Multiple types with normalization support

• Visualization: Real-time rendering with debug overlays (Frenet projection, lookahead curvatures, track lines)

See Installation to get started.

GETTING STARTED

• Installation
  • Using pip (recommended)
  • Using poetry
  • Additional dependencies
• Quickstart
  • Running examples
  • Environment basics
  • Drifting in one line
  • Simulation loop

USER GUIDE

• Configuration
  • Default configurations
  • Vehicle model and control
  • Training mode
  • Track direction
  • Observation configuration
  • Collision and safety
  • Reward configuration
  • Reset options
  • Debugging and visualization
  • Custom maps
• Training
  • Training scripts
  • Callbacks
  • Observation min/max snapshots
  • Instability prevention
  • log_std schedule
  • Curriculum learning
  • Wandb integration
  • Formatting and linting
• Sim2Real Transfer
  • Domain Randomization
• Architecture
  • Package structure
  • Important files
  • Vehicle parameters
  • Key dependencies
• Controllers
  • Pure Pursuit
  • Stanley and PD Controllers
  • MPC Controllers
• Analysis
  • Tire parameters
  • Analysis scripts
• Known Issues
  • Windows
  • macOS Big Sur and above

API REFERENCE

• Gym Environment
  • Creating an environment
  • Vehicle parameter methods
  • Step and reset
  • API reference
• Base Classes
  • RaceCar
  • Simulator
  • API reference
• Dynamic Models
  • KS — Kinematic Single Track
  • ST — Single Track
  • STP — Single Track Pacejka
  • STD — Single Track Drift
  • MB — Multi-Body
  • Vehicle parameters
  • API reference
• Observation System
  • Observation types
  • Configuration
  • Lookahead observations
  • Normalization
  • API reference
• Action System
  • Action space
  • Control input types
  • Configuration
  • Normalization
  • API reference
• Track System
  • Track loading
  • Map management
  • Frenet coordinates
  • Centerline and raceline
  • API reference

PROJECT

• Changelog