alvaro-frank/3d-bpp
A production-grade Deep Reinforcement Learning project solving the 3D Bin Packing Problem. It features DQN and PPO agents, a custom Gym environment, MLOps integration with MLflow and DVC, and a FastAPI service for inference.
What's novel
A production-grade Deep Reinforcement Learning project solving the 3D Bin Packing Problem. It features DQN and PPO agents, a custom Gym environment, MLOps integration with MLflow and DVC, and a FastAPI service for inference.
Code Analysis
15 files read · 4 roundsSolves the 3D Bin Packing Problem using DQN and PPO reinforcement learning agents in a custom Gym environment, with a BLB heuristic baseline, MLflow experiment tracking, and 3D visualization.
Strengths
Genuine RL implementations (PPO with GAE, DQN with target network) applied to a non-trivial spatial problem, with proper action masking, multi-component reward shaping, and a complete MLOps pipeline (Hydra config, MLflow, ONNX export, Docker, CI).
Weaknesses
Test coverage is minimal (only 7 basic unit tests for bin/box mechanics, no agent or integration tests), the heuristic baseline has a bug (iterates y over bin.height instead of bin.depth), and the reward function's compactness calculation uses unrotated box dimensions which is inconsistent with the rest of the code.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
116
Contributors
2
Files
35
Active weeks
17
Repository
Language
Python
Stars
3
Forks
1
License
MIT