Skip to content

PROJECT_STRUCTURE.md

Latest commit

 

History

History
103 lines (82 loc) · 3.06 KB

PROJECT_STRUCTURE.md

File metadata and controls

103 lines (82 loc) · 3.06 KB
Raw

Project Structure

Root Directory (Original Challenge Files)

controls_challenge/
├── README.md              # Original challenge README
├── requirements.txt       # Python dependencies
├── tinyphysics.py        # Core simulator
├── eval.py               # Evaluation script
├── controllers/          # Controller implementations
├── models/               # ONNX model files
├── data/                 # CSV route data
└── imgs/                 # Images for documentation

Our Additions

/analysis/ - Research & Analysis Scripts

All analysis, research notes, and investigative scripts:

  • BLACKBOARD.md - Research brainstorming
  • PHYSICS_ANALYSIS.md - Physics-based analysis
  • SPINNING_UP_LESSONS.md - PPO insights from OpenAI Spinning Up
  • beautiful_lander.py - Reference PPO implementation
  • verify_baseline.py - Baseline verification script
  • Other analysis scripts

/experiments/ - Experiment Implementations

Organized by experiment number (exp001, exp002, etc.):

  • exp017_baseline/ - 1-neuron PID clone (baseline)
  • exp023_conv/ - Conv1D on curvatures
  • exp030_vehicle_state/ - Vehicle-centric state representation
  • Each experiment has its own folder with:
    • train.py - Training script
    • controller.py - Controller for evaluation
    • model.pth - Trained model
    • README.md - Experiment notes

/archive/ - Old Attempts & Historical Code

Archived code from early attempts, kept for reference

/docs/ - Documentation

General project documentation

Current Best Results (Official Eval)

Baseline:
- PID: ~107 total_cost (1000 segs)
- exp023 (Conv BC): ~103 total_cost  ✅ Best BC
- exp030 (Vehicle state): ~112 total_cost

Target: <45 (56% improvement needed via PPO)

Key Files

Controllers (in /controllers/)

  • pid.py - Baseline PID controller
  • exp023_conv.py - Best BC controller (Conv1D)
  • exp030_vehicle.py - Vehicle-centric BC

Experiments (in /experiments/)

  • exp017_baseline/ - Proof that MLP can clone PID
  • exp023_conv/ - Conv1D breakthrough (12% gain over naive BC)
  • exp030_vehicle_state/ - Vehicle-centric representation

Running Experiments

Official Evaluation

# Test a controller
python tinyphysics.py --model_path ./models/tinyphysics.onnx \
                     --data_path ./data \
                     --num_segs 1000 \
                     --controller exp023_conv

# Compare against baseline
python eval.py --model_path ./models/tinyphysics.onnx \
              --data_path ./data \
              --num_segs 1000 \
              --test_controller exp023_conv \
              --baseline_controller pid

Training New Experiments

# Create new experiment folder
mkdir -p experiments/exp031_my_experiment

# Train
python experiments/exp031_my_experiment/train.py

# Test
python tinyphysics.py --controller exp031_my_experiment

Notes

  • Keep root clean - match original challenge structure
  • All research/analysis goes in /analysis/
  • All experiments go in /experiments/expXXX_name/
  • Archive old code in /archive/