Efficient and Real-Time Motion Planning for Robotics Using Projection-Based Optimization

Xuemin Chi, Hakan Girgin, Tobias Low, Yangyang, Xie, Teng Xue, Jihao Huang, Cheng Hu, Zhitao Liu, Sylvain Calinon

Zhejiang University, Hangzhou, China,

S3C robotics, Biel, Switzerland

Idiap & EPFL, Martigny, Switzerland

Code

Github

Paper

Under review.

Abstract

Generating motions for robots that interact with objects of various shapes is a complex challenge, further complicated when the robot's geometry and multiple desired behaviors are considered. Geometric aspects of robotic problems are often not fully exploited in inverse kinematics, collision avoidance and manipulation planning. To address this, we propose an efficient real-time approach that exploits analytical and learning-based geometric constraints via Augmented Lagrangian spectral projected gradient descent. Geometric constraints are designed and exploited via Euclidean projections, Minkowski sums and basis functions. We show that using geometric constraints instead of full constraints and gradients improves performance. We validate the efficiency of our through toy examples and extensive simulations, and provide comparisons against benchmarks, Additionally, we demonstrate its effectiveness on a 7-axis Franka robot and 1:10 scale car in real-world experiments.