tarzanagh/Multi-Robot-application

Multi-Robot Application

Code, trained weights, training data and measured results for the multi-robot motion-planning study (SMD / MMD / MPD / DM / EECBS / PCD / DiRecT / MDOC, plus our own certified planner) on the MMD and SMD benchmark suites.

Layout

path	contents
code/	all repos, no data, no .git
models/	model_current_state_dict.pth + ema_model_current_state_dict.pth + args.yaml per model (86 models)
data/	training demonstrations and benchmark test instances
results/	measured outputs: pickles, baselines_eval, EECBS
env/	conda specs -- rebuild with conda env create -f env/smd.yml

Continuing training on another GPU

conda env create -f env/smd.yml
conda activate smd
# point the trainer at data/data_trajectories and models/<model_id>/

args.yaml beside each checkpoint records the network shape (unet_input_dim, n_diffusion_steps, ...) needed to rebuild it.

What is deliberately absent

• Per-epoch checkpoint snapshots (62 GB). The trainer builds a fresh Adam on every call and never writes optimizer state, so restarting from the current state_dict loses exactly what restarting from a snapshot loses. They are useful only for branching from an earlier training point.
• Full-object *current.pth duplicates (3.9 GB). Same weights as the state_dicts, but pickled whole and fragile across torch/python versions.
• models_retrained (18 GB) -- retraining scratch.
• Conda environments (59 GB) -- rebuild from env/*.yml.

Total shipped: ~7 GB, from ~95 GB on disk.