DiPPeR: Diffusion-based 2D Path Planner
applied on Legged Robots

Jianwei Liu*             Maria Stamatopoulou*             Dimitrios Kanoulas

arXiv Video


Abstract

In this work, we present DiPPeR, a novel and fast 2D path planning framework for quadrupedal locomotion, leveraging diffusion-driven techniques. Our contributions include a scalable dataset generator for map images and corresponding trajectories, an image-conditioned diffusion planner for mobile robots, and a training/inference pipeline employing CNNs. We validate our approach in several mazes, as well as in real-world deployment scenarios on Boston Dynamic's Spot and Unitree's Go1 robots. DiPPeR performs on average 23 times faster for trajectory generation against both search based and data driven path planning algorithms with an average of 87% consistency in producing feasible paths of various length in maps of variable size, and obstacle structure.


In the following sections we present DiPPeRs performance and genarilisation capabilities.

Validation Dataset


We validate DiPPeR's performance on maps on the Validation Datase, which comprises of maps not previously encountered
during the training but share a similar structural pattern with those in the Training Dataset.


Short Trajectories

Path Length < 100

Medium Trajectories

100 < Path Length < 200

Long Trajectories

Path Length > 200



Out-of-Distribution Dataset


We evaluate DiPPeR's generalisation capabilities on maps that differ from those present in the Training Dataset.


OOD Dataset

Dataset consist of maps of different size, obstacle structure and colour, to test the generalisation capabilities of DiPPeR.

MRPB Dataset

Dataset used to compare DiPPeR's inference speed with that of A*, N-A* and ViT-A*. DiPPeR's inference time is 0.4s for all maps, regardless of their size, which is on average 23 times faster against the next best performing SOTA planners.




Real World Deployment

DiPPeR is integrated into the 2D ROS Navigation Stack, serving as the global path planner. When provided with the occupancy map, DiPPeR generates an initial global path. This path is subsequently fine-tuned by the TEB local planner to ensure the avoidance of violations of the robot's kinodynamic constraints. Additionally, the Phasespace Localisation system is employed to mitigate for state estimation inaccuracies.

BibTeX


      @article{liu2024dipper,
      title     = {DiPPeR: Diffusion-based 2D Path Planner applied on Legged Robots},
      author    = {Jianwei Liu, Maria Stamatopoulou, Dimitrios Kanoulas},
      journal   = {arXiv preprint arXiv:2309.14341},
      year      = {2024}
}