AI, vision & medical imaging4040 stars on GitHub
Brain MRI Pipeline
An end-to-end Nipype pipeline with FSL FAST, K-means and ANTs Atropos segmentation, tissue volumes, bias-correction metrics and an interactive 3D brain.
2018Completed
AI/ML
QWOP with PPO
Learning to finish, one knee-scoot at a time
A PyTorch PPO agent plays the real QWOP browser game through Playwright. Six evaluation races, six 100 m finishes. Grace was not part of the reward function.
- Role
- Creator & Maintainer
- Period
- 2019 – 2026.11
- Status
- Completed
- Theme
- AI, vision & medical imaging
Overview
Rebuilt in 2026 with Stable-Baselines3 PPO, PyTorch and a Gymnasium environment driving the official browser game through Playwright. Eight browser environments trained for 22 minutes. The recorded evaluation finished all six races; the best normal-speed game time was 193.8 seconds.
The problem
- Learning a stable gait from browser-game observations
- Measuring normal-speed game time consistently
What I did
- PPO with a Gymnasium environment and Playwright controls
- Fixed 60-frame game seconds and explicit evaluation records
Results
- Completed 100 m in all six recorded evaluation races
- Best normal-speed game time: 193.8 seconds
- Training run: 22 minutes across eight browser environments
Features
- Official QWOP game in headless Chromium
- Eight parallel training environments
- OCR distance feedback and shaped rewards
- Training curves, checkpoints and run GIFs
Related projects
AI, vision & medical imaging11 stars on GitHubStreet Scene OCR
Compare EasyOCR, RapidOCR and PaddleOCR on Traditional Chinese street signs, with text boxes, timings and a small public-domain benchmark.
2026Completed
AI, vision & medical imaging11 stars on GitHubSentiment Bundle Benchmark
TF-IDF, embeddings, fine-tuned encoders and a local language model, evaluated on fixed splits across six sentiment datasets. The successor to GURA.
2026Completed