Robots are being used these days to perform various tasks and their algorithm is important for them to perform the actual work. Researchers at UC Berkeley have developed a RoVi-Aug framework which will augment robotic data and will help transferring the data in other robots. Modern machine learning and generative models have generalized their data which is used in almost all of the models. So, the researchers wanted to make something similar to that for the robots that can generalize their data.
The researchers had been trying to generalize robotic data since the start of this year and had been doing various experiments on it. In their previous research, the researchers realized that there are some challenges in generalizing robotic data too. They found that if the robotic data is unevenly distributed, it can become less effective in teaching other robots the same skills.
But the researchers soon found out that a lot of robots have uneven datasets, including the Open-X Embodiment (OXE) dataset which is widely used for training robotics algorithms. This type of imbalance can limit the performance of robots too. To solve this issue, researchers proposed a new algorithm called Mirage which uses a technique called cross-painting to transform unseen robots into source robots. But there are some limitations of this algorithm too.
First of all, it needs exact robot models and cameras, and cannot adjust with camera angles. As an alternative, the researchers presented RoVi-Aug which is flexible and adaptable and can create synthetic images that show robots tasks from different angles.
RoVi-Aug also doesn’t require any extra processing during its deployment and allows changing camera angles from different perspectives. It can help researchers to train other robots because it has precise camera setups which are essential for robot training. RoVi-Aug is also cost-effective and can help other robots improve in learning and training.
Image: DIW-Aigen
Read next: New Study Shows Parents Prefer AI for Child Healthcare Advice, Raising Concerns
The researchers had been trying to generalize robotic data since the start of this year and had been doing various experiments on it. In their previous research, the researchers realized that there are some challenges in generalizing robotic data too. They found that if the robotic data is unevenly distributed, it can become less effective in teaching other robots the same skills.
But the researchers soon found out that a lot of robots have uneven datasets, including the Open-X Embodiment (OXE) dataset which is widely used for training robotics algorithms. This type of imbalance can limit the performance of robots too. To solve this issue, researchers proposed a new algorithm called Mirage which uses a technique called cross-painting to transform unseen robots into source robots. But there are some limitations of this algorithm too.
First of all, it needs exact robot models and cameras, and cannot adjust with camera angles. As an alternative, the researchers presented RoVi-Aug which is flexible and adaptable and can create synthetic images that show robots tasks from different angles.
RoVi-Aug also doesn’t require any extra processing during its deployment and allows changing camera angles from different perspectives. It can help researchers to train other robots because it has precise camera setups which are essential for robot training. RoVi-Aug is also cost-effective and can help other robots improve in learning and training.
Read next: New Study Shows Parents Prefer AI for Child Healthcare Advice, Raising Concerns
