Model validation

The probability modelP(X = 1|θ)⃗ in Section 3.8.2 wasfitted using the sampling procedure in Section 3.8.3. For all tasks approximately 3,300 valid samples were generated around task canonical poses. The estimated probability models were val-idated by sampling each dimension ofθ⃗separately on grid points and executing the task ten times on each point with real robot. The averaged task success rate on real robot was then compared against the proposed models and the estimated probabili-ties matched well as can be seen in Fig. 3.12.

Next, the proposed metric was compared against the ADC metric in controlled experiments. For each task we generated a synthetic set{Υˆ_i}^N_i=1of 6D poses where each poseΥˆ_idiffers from the ground truth poseΥ=I either by rotation or transla-tion along a single axis. The success probability was estimated using the procedure in Sec. 3.8.4 and the ADC score as described in Sec. 3.8.1. The results are shown in Fig. 3.13. It is important to notice that the ADC error cannot take into account the intrinsic parameters of the tasks and assigns the the same performance score for the hardest (Task 1) and easiest task (Task 4) over the same set of object poses. In contrast, we can see that our proposed performance metric can clearly indicate that Task 1 has significantly less tolerance in the pose estimation errors and thus is much harder to conduct. Moreover, the proposed metric can measure the turning point after which the success probability drops quickly from 1.0 to 0.0 where the as ADC error regrades linearly even after these points and is thus uninformative. The dif-ference between the two metrics is further illustrated in the success probability vs.

ADC scatter plots of all four tasks in Fig. 3.14.

3.9 Summary

In this chapter, the 6D pose estimation pipeline was described along with the con-tributions. In specifically, we focused on point cloud and correspondence-based ap-proaches, where point pair matches between two input objects are predicted based on local support, i.e. based on geometric surface properties of a small region. Finally, a pose (transformation matrix) is estimated that aligns the corresponding point pairs based on some distance function. Usually the pose search is guided by exploiting ge-ometric constrains between the correspondences and using robust techniques such

Figure 3.12 Engine cap used in our experiments. The coordinate system is object centric (top left) and pose samples are taken around a canonical grasp pose. Below are the estimated (the red, green and blue lines) and validated success probabilities (yellow line) on the six main axes (three translations and three rotations) in vicinity of the canonical grasp pose.

as the well known RANSAC or point clustering.

During the chapter, we descried two alternative methods to improve the perfor-mance of a pose estimation pipeline: curvaturefilteringandregion pruning. In our preliminary experiments, the robustifying methods were able to identify a robust sub-set of points against estimation failures and consistently improved the three cor-respondence based methods: GC[21], HG[127]and SI[13]. However, there was no clear winner between the two robustifying methods and a new study was conducted with larger dataset. The main result on the new dataset was that the GC with opti-mized parameters outperformed all the other methods on most of the object classes.

However, unlike in our previous work with limited data, the two robustifying ap-proaches did not provide systematic improvement after the meta parameters of the pose estimation methods were tuned.

In thefinal section of the chapter, we discussed different techniques for measuring the estimated object pose and introduced completely new metric for robotic

manip-Figure 3.13 ADC and success probability from controlled experiments for all the tasks (Task 1 – Task 4). Effect of rotation (left column) and translation (right column) to the ADC and success probability.

ulation. In our experiments we demonstrated how the popular error measure, ADC, poorly indicates success in robot manipulation tasks and is therefore uninformative.

As a novel solution, we proposed a probabilistic metric that measures the true suc-cess rate without the physical setup and provides basis for more realistic evaluation

Figure 3.14 Success Probability vs. ADC scatterplot from controlled experiments for all the tasks (Task 1 – Task 4). The scatterplot shows that the ADC does not reflect the success probability, except for extreme and trivial cases of failure or success; the two measures cannot be put in correspondence to each other not even through a nonlinear mapping.

of object pose estimation methods.

4 SAFE HRC IN INDUSTRIAL MANUFACTURING

4.1 Introduction

For decades, industrial robots have been irreplaceable resource for manufacturers, being efficient in repeatable and simple tasks. In isolated workcells the robots can operate without any external sensors and apply simple strategies to succeed in tasks, such as in welding and part feeding. However, demand for moreflexible and collab-orative systems is rising and currently the industrial manufacturing is going towards a new industrial revolution, the so-calledIndustry 4.0. Human-robot collaboration (HRC) will have an important role in the shift and this evolution means breaking with the established safety procedures as the separation of workspaces between robot and human operator is removed. However, this will require special care for human safety as the existing industrial standards and practices are based on the principle that operator and robot workspaces are separated and violations between them are monitored.

HRC has been active in the past to realize the future manufacturing expectations and made possible by several research results obtained during the pastfive to ten years within the robotics and automation scientific communities[140]. In particular, this has involved novel mechanical designs of lightweight manipulators, such as the Uni-versal Robot family¹and KUKA LBR IIWA². Due to the lightweight structure, slow speed, internal safety functions and impact detection, the robots are considered a more safe solution for close proximity work than traditional industrial robots. The collaborative robots can be inherently safe, but the robotic task can create safety haz-ards for instance by including sharp or heavy objects that are carried at high speed.

1https://www.universal-robots.com/

2https://www.kuka.com/en-de/products/robot-systems/industrial-robots/lbr-iiwa

In order to guarantee the safety of the human co-worker, a large variety of external multi-modal sensors (camera, laser, structured light etc.) have been introduced and used in robotics applications to prevent collisions[51, 110]. Moreover, in order to transfer research solutions from the research lab to real industrial environments they need to comply with strict safety standards.