Long-horizon Planning with Hybrid Diffusion

Implementation of planning with hybrid diffusion. Includes code for training and evaluation of the method.

Reproducing experiments

To reproduce the experiments, do the following:

1. Collect demonstrations using the scripted expert
2. Train and evaluate policies

Collecting demonstrations using scripted expert

The demo-collection scripts are provided in hybrid_diffusion_planning/dataset/, with the following table listing which script corresponds to which task.

Task	Demo Collection Script
Tool-use	collect_hook_demos.py
Arrange	collect_xarm_arrange_demos.py
Xarm Sorting	collect_xarm_sorting_demos.py
2D Sorting	collect_sorting_demos.py

Single-seed Training

For running training and evaluation, run the associated config file. For HDP for the Hook-task, run

python train.py --config-name=train_hybrid_xarm_hook

Multi-seed training

We use the evaluation functionality from Diffusion Policy to obtain metrics averaged across 10 epochs and 3 seeds. The multi-seed training is initiated by:

export CUDA_VISIBLE_DEVICES=0,1,2
ray start --head --num-gpus=3  
python ray_train_multirun.py --config-dir=diffusion_policy/config --config-name=$config --seeds=42,43,44 --monitor_key=test/mean_score

Metrics are logged continually during training, and we report the k_min_train_loss metric in the paper.

See the Diffusion Policy repo for more information.

Acknowledgements

This codebase is based on Diffusion Policy.