Improvement in reactiveness of realtime trajectory smoothing

Abstract

Human-robot collaboration has been an advancing topic throughout the years. Therefore for human-robot collaboration to work safely and efficiently, a real-time system must be fully developed. One important part of the system is the Realtime Motion Planning (RMP). RMP requires extremely fast computation throughout the whole process from detecting obstacles to planning of trajectories. However, by just being able to do RMP is not enough as with RMP only the path executed will be jerky.

A Realtime Trajectory Smoother (RTS) was proposed to solve the issue of jerky trajectory in RMP. With the RTS, it can perform obstacle avoidance while executing a smooth trajectory. Within the RTS is one of its crucial parts which is the implementation of Neural Network Based (NN-Based) Collision Estimation by using a Clearance Field Neural Network (CFN). We propose to replace CFN with a more direct and easier processing method through either Voxel Precomputation Based (VP-Based) Parallel Collision Detection or Signed Distance Field Based (SDF-Based) Collision Detection as possible alternatives for the system. The proposed alternatives bring mainly 2 advantages, which is the removal of neural network as well as removing posterior geometric trajectory verification.

SDF Based Parallel Collision Detection takes fairly less preparation time but in turn takes up more time during operation. VP-Based Parallel Collision Detection on the other hand takes more preparation time for collision status precomputation but produces fast results during operations, which is far more feasible in our case.

Furthermore, this research aims to achieve real-time collision detection when implemented. At the end of the report, we will conclude our report by discussing about the different improvements and optimizations available for future references.