Abuudiii/qrdqn-QWOP

QR-DQN agent finishing QWOP's 100 m

A Quantile Regression DQN agent (sb3-contrib) trained on qwop-python, a headless Box2D recreation of Bennett Foddy's QWOP.

• Result: finishes the 100 m (100.7 m in 107.1 s of game time), deterministic policy
• Training: 10M steps, 8 parallel games, 80 min on an AMD Radeon AI PRO R9700 (ROCm)
• Checkpoint: 9.0M steps, the best of 20 checkpoints (12 of them finish the race)
• Actions: 9 Q/W/O/P key combinations, each held for 4 frames
• Observations: 60 numbers (body-part positions, angles and velocities)

Settings: train_qrdqn_10m.yml (the qwop-python world-record attempt config train_qrdqn_wr.yml, shortened to 10M steps and run with 8 parallel games) and env.yml.

Code: https://github.com/Abuudiii/rl-experiments

Usage

from sb3_contrib import QRDQN
model = QRDQN.load("qrdqn-QWOP.zip")