Title: 5-axis path planning with fast collision detection and deep reinforcement learning

Xin Chen

Ph.D. Student

School of Computer Science

College of Computing

Georgia Institute of Technology

Date: Tuesday, January 15th 2019
Time: 12:00pm – 1:30pm (EDT)
Location: KACB 1315

Committee:
----------------
Dr. Richard W. Vuduc (Advisor, School of Computational Science and Engineering, Georgia Institute of Technology),

Dr. Ümit V. Çatalyürek (School of Computational Science and Engineering, Georgia Institute of Technology)

Dr. Thomas R. Kurfess (School of Mechanical Engineering, Georgia Institute of Technology)

Abstract:
----------------

5-axis machining is a strategy that allows CNC move an object or cutting tool along five different axes (X, Y, Z and two additional rotary axes) simultaneously. This provides infinite possibilities of machining very complex objects, which is why 5-axis machining gets more and more popular. This thesis focuses on a path planning problem that arises in 5-axis machining applications: how to construct a tool path that covers the surface of a 3D object, produces a short milling time, and is collision-free. This thesis proposes a deep learning framework with a fast collision detection algorithm to generate an efficient 5-axis path.

We first present a collision detection algorithm, named aggressive inaccessible cone angle(AICA) for CNC milling to accelerate the intersection test. The key idea of our proposed method is the concept of inaccessible cone angle(ICA), which is a new geometric abstraction for collision detection problem, and its effective use, including memoization to remove redundant work and increasing parallelization. We have prototyped our AICA algorithm within a real CNC milling tool, SculptPrint. Experimental results on 4 CAD benchmarks demonstrate that AICA is up to 23 times faster than the approach of the traditional checking.

Second, this thesis presents a new path planning algorithm with deep reinforcement learning, called Deep path, to generate an efficient path for an arbitrary 2D environment. The key idea of this algorithm is a new graph model based on Boustrophedon Cellular Decomposition (BCD), which is a method of transforming a space into cell regions with morse decomposition. This graph model can easily reflect the physical distance in the graph, and evaluate the cost of an arbitrary path. We show that when applied into deep reinforcement learning, Deepath can efficiently reduce the path length with a complete coverage on the environment.

Third, this thesis proposes a new 5-axis path planning algorithm with deep reinforcement learning, considering both the trajectory of the cutting tool end in 3-axis, and the orientations of the tool as the other 2 rotatory axes to guarantee a collision-free path. This algorithm aims at reducing the machining time, by designing an efficient 5-axis path to avoid tool retractions (pulling the tool back and in).