Improving Robot Integration in Small Businesses

In today's world, robots are becoming more common in various industries. However, many small and medium-sized companies still struggle to integrate robots into their workflows due to the need for constant adjustments when the environment changes. This can make automating tasks difficult, especially tasks that are usually done by humans, such as assembling products. One way to help robots adapt to their surroundings is by creating a model of the environment, using data collected from a 3D camera.

#The Challenge

Many companies face challenges when trying to use robots for tasks like assembly. When the workspace changes, robots often need their programs updated, which can be time-consuming. This is especially true for smaller businesses that lack the resources to make these adjustments frequently. When robots cannot adapt quickly, it leads to wasted time and makes the robots less efficient.

#The Solution

Creating a Semantic Model from Point Clouds, which is a collection of data points in space, can help robots better understand their surroundings. By segmenting these point clouds into parts that the robot can recognize, they can automate tasks more effectively.

There is a gap in research when it comes to applying semantic segmentation to point clouds for industrial robots. This paper discusses how to create artificial point clouds to train models that can better segment these clouds. By using our method, we can train models that reach high accuracy in real-world scenarios.

#Methodology

#Dataset Creation

Since it is hard to gather real-world data, we decided to make most of our dataset using synthetic data. Our dataset has 2,218 labeled point clouds, with 90% being synthetic and 10% real. Four types of objects are included: storage racks, boxes, shadow boards, and tables.

To generate labeled point clouds, we used a tool that allows for changes in the scene, such as altering the location, size, and appearance of objects. Using software like Blender, we can create various scenes and capture different object configurations.

#Exploring the Workspace

When using a robot in an unfamiliar environment, the user teaches the robot where to capture point clouds from. The robot then takes these captures to create a 3D model of the space. Each point cloud from the captures is then processed using a Segmentation Network to classify the objects.

#Segmentation Processing

Once we have the labeled point clouds, we analyze them to determine where the robot will move and how it will grasp objects. First, we find all the boxes in the point cloud and identify individual boxes using a clustering method. Then we look for the shadow board, which helps to determine where the robot will place objects.

To fit the shadow board into the point cloud accurately, we use a method that finds features in both the CAD model and the captured point cloud. This helps to match the two, allowing the robot to know exactly where to place the objects.

#Assembly Execution

During the assembly process, the robot will grasp objects and place them on the shadow board. While the grasp planning and placement are areas for further research, our current focus is on developing a reliable method for finding objects and determining their positions.

#Results and Validation

#Workspace Exploration

We conducted experiments to see how well our trained models could navigate and understand the workspace. We tested different model architectures and found that models trained with a mix of real and synthetic data performed better than those trained only with synthetic data.

Our results indicated a significant gap between synthetic and real-world performance, meaning that adding real-world data made the models more effective. When using different cameras, our approach maintained accuracy across various camera types, proving its robustness.

#Shadow Board Registration

Finding the shadow board in the workspace is crucial, as even minor errors can disrupt the assembly process. We evaluated our registration algorithm against point clouds from different cameras to see how well the shadow board could be fitted into a real-world point cloud.

It became clear that a high-quality point cloud is essential for accurately registering the shadow board. Lower-quality cameras struggled to provide reliable results, indicating the need for stronger registration methods.

#Conclusion and Future Work

This research presented a method for helping robots understand their environment with minimal human help. By creating a dataset that combines both synthetic and real-world data, we've trained models that can accurately identify objects in a point cloud.

The challenge of adapting robots to different environments is easier with our approach, which emphasizes creating a semantic world model. In the future, we plan to study more robust registration techniques to aid the fitting process and reduce our reliance on high-quality point clouds. Additionally, we aim to implement strategies that allow for more autonomous exploration, which can further decrease the need for user intervention.

Our findings indicate that automation in small and medium-sized enterprises is within reach, making it possible for these companies to embrace the advancements in robotic technology. By continuing to simplify the integration of robots into everyday tasks, we can help pave the way for a more efficient future in the workplace.