This repository is an academic research fork of the official NavDP implementation. The purpose of this fork is to deepen the understanding of Diffusion Policies in robotics and extend the framework's capabilities through custom hardware integration and enhanced visualization tools.
In this sandbox, I have implemented several critical enhancements to the core NavDP framework:
Successfully integrated the LeKiwi robot model, transitioning from the default differential-drive Dingo to a more complex omni-directional 3-wheeled base.
Technical Resources:
- LinkedIn Article: The Ghost, The Anchor, and The Exploding Robot: A LeKiwi Integration Log
- Repo Integration Log: LEKIWI_INTEGRATION_LOG.md — A deep dive into why off-the-shelf USD models are often "simulation-broken" and how to fix them for navigation.
The integration required significant engineering efforts in:
- USD Preparation: Programmatically fixing articulation roots and programmatically injecting collision geometries.
- Physics Tuning: Solving simulation "explosions" (NaNs) by optimizing actuator gains and contact offsets.
- Friction Modeling: Implementing a hybrid collision model (Cylinders for traction wheels, frictionless Sphere for the caster) to ensure realistic motion and turn-in-place capabilities.
Expanded the visual evaluation capabilities by adding two primary research perspectives:
- Bird's Eye View (BEV): For clear path-planning analysis and spatial reasoning.
- Third-Person View: To better observe the robot's physical interaction with the environment.
Developed a suite of diagnostic and integration tools located in the research_tools/ directory to automate the robot on-boarding process and physics verification.
To maintain a research-ready environment, this fork follows a structured hierarchy:
research_tools/: Automation scripts for USD fixes, collision generation, and physics validation.evaluation_outputs/:benchmark_runs/: Raw data from evaluation episodes.videos/: Recorded simulation runs showcasing navigation behaviors.
development_logs/: Detailed troubleshooting logs (e.g.,lekiwi_integration_log.txt) documenting the technical journey.README_OFFICIAL.md: The original documentation with full technical details of the NavDP paper.
Follow the installation steps in README_OFFICIAL.md to set up the environment.
To run evaluations with the new LeKiwi configuration:
# Example PointGoal evaluation with LeKiwi
python eval_pointgoal_wheeled.py --port 8888 --robot lekiwi --scene_dir [PATH_TO_SCENE]This work is built upon the NavDP (Navigation Diffusion Policy) paper:
NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance
Wenzhe Cai, Jiaqi Peng, Yuqiang Yang, Yujian Zhang, Meng Wei, Hanqing Wang, Yilun Chen, Tai Wang and Jiangmiao Pang
Official Repository | arXiv Paper
I express my gratitude to the original authors for open-sourcing their impressive work, which provided the foundation for this research exploration.