How charlesxu0124/cablerouting works
Project type and purpose
• This is a research-oriented machine learning pipeline (“CableRouting”) implementing a multi-stage cable routing policy via hierarchical imitation learning.
• It ingests geometric trajectory data, preprocesses it, pretrains a ResNet encoder, learns a low-level routing policy (behavior cloning), learns a high-level primitive-selection policy, then fine-tunes end-to-end. Logging and checkpointing integrate with W&B.
File structure
• Top-level directories:
• data/ – raw and processed datasets; contains processing script process_route_trajectories.py
• local_scripts/ – bash scripts orchestrating each training stage
• src/ – core Python code: entry points (bc_main.py, primitive_selection_main.py), model definitions (model.py), utilities (data.py, jax_utils.py, train_utils.py, utils.py)
• requirements.txt – dependency list (jax, torch, wandb, etc.)
• README.md – describes data download, environment setup, and training workflow
README for additional insights
• Four main training stages, each invoked via a script:
1) pretrain_resnet_embedding
2) train_routing_bc
3) train_highlevel
4) finetune_highlevel
• Uses JAX for embedding pretraining, PyTorch for policies, and W&B for logging.
Detail
• Every stage as a distinct component
• Environment setup (conda env, requirements.txt) as a preliminary step or footnote
• Don’t omit any data transformation or checkpoint artifact