How shirtep/fly-mania works
fly-mania trains a fly-inspired controller to play osu!mania 4K, then uses that same inference path for evaluation and video rendering. The README documents the perception-to-action workflow, fixed connectome, training-only DAgger teacher, physical actuator, and official scoring. The graph also includes the major rendering and training subsystems. Some connections rely on README or SOURCE INDEX references; unsampled implementation details are not inferred.
Main entry points: video_renderer.py Video renderer · train.py Training entry · play_map.py Map simulation · policy_runtime.py Policy runtime
Components 23
Training
- Rollout packs
src/training/rollout_packs.py - DAgger teacher
src/training/teacher.py - Rollout learning
src/training/rollout_loss.py - Training entry
scripts/train.py
Gameplay
- Causal perception
src/environment/causal_map.py - Fixed connectome
src/training/connectome.py - Recurrent policy
src/training/policy.py - Physical actuator
src/environment/actuator.py - Policy runtime
src/training/policy_runtime.py - Official scoring
src/environment/scoring.py - Map simulation
src/evaluation/play_map.py
Rendering
- Render interface
scripts/render_ui.py - Video renderer
src/viewer/video_renderer.py - 3D scene
src/viewer/scene_renderer_3d.py - Fly kinematics
src/viewer/fly_kinematics.py - Connectome HUD
src/viewer/connectome_hud.py - Neural display
src/viewer/neural_display.py - Game screen
src/environment/mania_screen.py - Text fonts
src/fonts.py
Maps and data
- Training corpus
src/training/corpus.py - Beatmap parser
src/environment/osu_parser.py
Other components
- Trainer
- Map user
Connections 27
- Trainer → Training entry: starts training
README.md - Training entry → Training corpus: loads corpus
scripts/train.py - Training corpus → Beatmap parser: parses maps
src/training/corpus.py - Trainer → Map user
- Map user → Render interface: selects map
README.md - Render interface → Video renderer: requests render
scripts/render_ui.py - Beatmap parser → Causal perception: provides beatmap
src/environment/causal_map.py - Training entry → Rollout packs: builds rollouts
scripts/train.py - Rollout packs → Causal perception: uses chart
src/training/rollout_packs.py - Training entry → DAgger teacher: trains against
scripts/train.py - DAgger teacher → Physical actuator: plans key actions
src/training/teacher.py - Training entry → Rollout learning: optimizes with
scripts/train.py - Rollout learning → Policy runtime: advances policy
src/training/rollout_loss.py - Policy runtime → Physical actuator: commands
src/training/policy_runtime.py - Policy runtime → Recurrent policy: runs inference
- Policy runtime → Fixed connectome: updates neural state
- Map simulation → Beatmap parser: parses map
src/evaluation/play_map.py - Map simulation → Causal perception: builds perception
src/evaluation/play_map.py - Map simulation → Policy runtime: advances policy
src/evaluation/play_map.py - Map simulation → Official scoring: judges events
src/evaluation/play_map.py - Video renderer → 3D scene: renders scene
src/viewer/video_renderer.py - 3D scene → Fly kinematics: uses motion
src/viewer/scene_renderer_3d.py - Video renderer → Connectome HUD: draws overlay
src/viewer/video_renderer.py - Video renderer → Neural display: draws activity
src/viewer/video_renderer.py - Video renderer → Game screen: draws game screen
src/viewer/video_renderer.py - Connectome HUD → Text fonts: loads text
src/viewer/connectome_hud.py - Video renderer → Map simulation: captures simulation
src/viewer/video_renderer.py