A Constrained Motion Planning Method Exploiting Learned Latent Space
for High-dimensional State and Constraint Spaces
Abstract
Novel approach for high-dimensional constrained motion planning
Introduces tangent space dataset augmentation to approximate manifold with sparse dataset
Introduces latent motion for efficient learned latent space search
and latent jump to mitigate topological problems.
Problem
Three joint robot with a positional constraint
Truth manifold
S-VAE
N-VAE
latent color map
(N-VAE)
Core Idea
Latent space search by alleviating manifold mismatch problem
Dataset augmentation for sparse constrained motion dataset
Resources
Source codes: https://github.com/psh117/ljcmp
Training and test datasets: Download
Pretrained model: DownloadÂ
Experiment Video
Learned Latent Space
This method utilizes the learned latent space with condition variables.
Right animation illustrates the constrained motion according to the latent code and conditions.
ROS Interface Demo
Real-time inverse kinematics and constrained motion planning for the pose of the interactive marker.
Motion planning is performed when the interactive marker is released, and the trajectory is displayed once, corresponding to the changed target pose.
Target tasks are dual arm manipulation with orientation constraint on the tray (parallel to the ground).
All videos are in real-time (1x speed) with no planning time skipped.
Demo 1 - Franka Panda
Demo 2 - Franka Panda
Demo 3 - DYROS RED Humanoid Robot
Demo 4 - DYROS RED Humanoid Robot
Demo 5 - Baxter
