Plan First, Diffuse Later: Extrinsic Graph Guidance for Long-Horizon Diffusion Planning

This is the official implementation for "Plan First, Diffuse Later: Extrinsic Graph Guidance for Long-Horizon Diffusion Planning".

Yaniv Hassidof¹, Adir Morgan¹, Yilun Du², Kiril Solovey¹

¹Technion, ²Harvard

---

🛠️ Installation

1. Initial Setup: Follow the CompDiffuser instructions for the base diffuser implementation and checkpoints.
2. XDiffuser: Copy the contents of XDiffuser.zip into the root folder.
3. Graph-Assisted Stitching (GAS):
   • Clone and install the GAS repository into diffuser/Roadmaps.
   • Download the appropriate checkpoints to the /checkpoints folder.
4. Bridge Inspection (Optional): For the bridge inspection experiments, the MILP solver is required. Follow instructions at ScalableInspectionPlanning.

---

🕹️ Experiments

OGBench Main Experiments

To launch the main OGBench experiments, toggle the appropriate config and run:

sh diffuser/Roadmaps/prm_ogb/plan_prm_dfu_ogb.sh

Multi-Agent Experiments

To launch multi-agent experiments:

sh diffuser/Roadmaps/prm_ogb_multi_agent/plan_prm_ogb_multi_agent.sh

The configuration for these experiments is located at: config/baselines/prm_ogb/prm_dfu_ogb_antM_Me_Multi_Agent.py

Inspection Planning Experiments

For bridge inspection experiments, directly run the script:

python diffuser/Roadmaps/prm_ogb_inspection_planning/plan_inspection_bridge.py

The configuration for these experiments is located at: config/baselines/prm_ogb/prm_dfu_ogb_bridgeQuad_Inspection.py

---

🙏 Acknowledgement

Our implementation is based on CompDiffuser and Graph-Assisted Stitching (GAS).

---

📝 License

This repository is released under the MIT license. See LICENSE for additional details.