Automatic annotation of accessibility data in a terrain to traverse a four legged robot

Freitag, 24. Juni 2022

An overview of the problem:

The earliest robot was made in the 1950s. Since then, there have been many developments in its mechanics to make it functional in many service areas. It was guided by a human all the time to make decisions. However, with the resurgence of Machine learning, as we improved our capabilities in computation, Robots are now developed to make some of the decisions on their own by recognizing pattern in its sensor data inflow. This task is part of a bigger development. An IMU mounted on a robot takes data while it moves through a certain area. We want to see if the area the robot is passing through is indeed accessible based on a statistical model. The basic aim for this specific work is to create a statistical model to (semi-)automatically annotate accessibility data to parts of map of a terrain. This annotated data will then be used by a four-legged robot SPOT mini to navigate in that terrain. There are two possible outcomes if the robot encounters area not accessible enough:

• Slippage

• Immobilization

solution concept:

The task is to create some rules for the robots to decide the terrain’s accessibility. This is why, we will put an IMU sensor in the robot and we will make it move through multiple terrains to get data and behaviour. As we do not want the robot to fall a lot, therefore we would like to develop this approach in a simulated environment. To simulate a robot, we will be using existing frameworks like ROS to supply external commands to the robot and gazebo to simulate the robot in an environment. This way, we will collect data from a ROS + Gazebo environment.

These data will then be analysed based on their behaviour primarily using python to find some rules that allow us to infer accessibility information. The analysis may contain but not limited to simple statistical models like classification, some explainable Machine learning techniques like support vector machine, random forest or linear regression. There are already some developments using self-organizing maps, K-means clustering and auto encoding. We will first implement these findings and then approach this problem in other ways