TeoIlie/Gym-Khana
Custom gym environment for training Deep Reinforcement Learning policies to race and drift on 1/10 scale or full-size Ackermann vehicles. SB3 and wandb integration. Based on the f1tenth_gym simulator built by UPenn
What's novel
Custom gym environment for training Deep Reinforcement Learning policies to race and drift on 1/10 scale or full-size Ackermann vehicles. SB3 and wandb integration. Based on the f1tenth_gym simulator built by UPenn
Code Analysis
8 files read · 3 roundsA Gymnasium RL environment for training racing and drifting policies on Ackermann-steered vehicles, featuring multiple physics models (kinematic, single-track, single-track-drift with Pacejka tires, and a 29-state multi-body model), laser scan simulation, Frenet coordinate tracking, and domain rando
Strengths
Genuinely deep physics implementations with Numba-optimized dynamics (Pacejka Magic Formula tire model, load transfer, wheel dynamics, smooth kinematic/dynamic blending), clean separation of concerns across action/observation/track/dynamics modules, and thoughtful engineering details like steering delay buffers, instability detection with state reversion, and ZOH-exact steering rate control.
Weaknesses
The multi-body model is not Numba-jitted (unlike ST/STD), creating a performance inconsistency; the main environment file is 59KB with significant coupling; and the observation module mixes Numba-optimized and pure-Python paths with some redundant validation logic.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
337
Contributors
13
Files
198
Active weeks
24
Repository
Language
Python
Stars
1
Forks
0
License
MIT