Asymptotically optimal path planning with integrated local optimisation for robotic vine pruning.

University of Canterbury Research Repository (University of Canterbury)2017

Abstract

Grapes are an important crop to New Zealand and the rest of the world. Vineyards cover 75,000km2 worldwide and New Zealand wine exports were valued at $1.57 billion in 2016. Grape vines are pruned annually to improve yield. Pruning is done manually for many vine species and is the most labour intensive and expensive task for the vineyard. A robotic pruning system is being developed at the University of Canterbury to automate cane pruning, where canes are selectively removed to leave a small number of long healthy canes. The robot uses stereo-cameras to construct a 3D model of the vines and an AI system to determine where vines should be pruned. A UR5 robot arm with a spinning router is then used to make the cuts. An online path planning algorithm is required to plan collision free paths for the robot arm to reach cut-points. The path planner should quickly compute paths that are fast to execute after being converted to trajectories so that the robot can operate efficiently. This thesis proposes new path planning approaches for quickly computing paths that are fast to execute when converted to trajectories. Sampling based path planners are by far the most widely used path planning algorithms for high degree of freedom robot arms. These planners explore the robot’s configuration space to find collision free path for the robot to follow. Some of these planners also attempt to optimise a solution as time permits. The performance of sampling-based path planners can be limited by the efficiency of the collision detector that they use, their (sometimes slow) convergence rate and how they can exploit a robot arm’s redundancy to find high quality paths. Efficient operation of the vine pruning robot relies on its path planner being able to quickly find paths that can be converted to fast-to-execute trajectories. Collision detection speed is often the bottleneck in performance for many sampling based path planners. Fast path planning can be achieved by using an efficient collision detector. This research proposes a new collision detection algorithm that exploits the structure of grape vines to provide fast collision detection. This collision detector takes 3.0 × 10−6 seconds on average to classify the collision status of a robot configuration and is 50 times faster than the popular Flexible Collision Library [Pan et al 2012]. This speed-up in collision detection results in a 27 times reduction in path planning times. Two approaches to finding fast-to-execute paths are to use an asymptotically optimal path planner, or to use a local optimiser. Asymptotically optimal planners are guaranteed to eventually find the optimal, e.g. shortest, solution but can be slow. Local optimisers are capable of quickly improving a solution with respect to a cost function such as path length, but are not guaranteed to find the optimal solution. The approach proposed in this thesis integrates an asymptotically optimal planner with a local optimiser to speed up the search for short paths while retaining the planner’s asymptotic optimality. The asymptotically optimal RRTConnect* planner integrated with a ‘short cut’ local optimiser found paths that were 31% faster to execute than those found by RRTConnect* without the local optimiser for the vine pruning robot given three seconds of planning time. Many robot arms are redundant with respect to their tasks. The robot arm might be able to accomplish the task, e.g. move the end-effector to a specific Cartesian position, using more than one set of joint angles. Ideally the robot’s path planner would be able to use the extra configurations to find higher quality paths, however, little work has been done to investigate this. In this thesis these extra goal configurations are used to find significantly shorter paths that are faster to execute compared to a planner that chooses one goal configuration arbitrarily. A planner using these redundant goal configurations found paths that had 58% lower execution times on average compared to a planner that did not use the redundant goal configurations for the vine pruning robot. This thesis investigates ways to reduce the computation time and improve the solution quality of sampling based path planners, specifically for the task of pruning grape vines using a robot arm. Fast computation times are achieved using a new collision detector that exploits the structure of grape vines, and by integrating an asymptotically optimal path planner with a local optimiser. The robot arm’s redundancy is exploited to allow the planner to discover fast-to-execute paths.

Abstract

Related Papers