Machines have been built to handle some of the toughest terrains, but for robots, learning the ropes is a little more difficult. State-of-the-art algorithms have provided robotic machines with the capability to navigate up hillsides, over rocks, and fallen branches. The main issue surrounding robotics is that artificial intelligence (AI) driven robots become easily confused by new environments.
In a paper, “Virtual-to-Real-World Transfer Learning for Robots on Wilderness Trails,” scholars at the University of Colorado believe they may have found a solution. The key lies in numerous deep learning models, such as layers of neuron-simulating mathematical functions, which examine camera footage to help the robot maneuver hiking trails.
UC researchers say, “Robots hold promise in many scenarios involving outdoor use, such as search-and-rescue, wildlife management, and collecting data to improvement environment; climate, and weather forecasting.” Outdoor trails hold a lot of variables, particularly when it comes to obstacles. “However, autonomous navigation of outdoor trails remains a challenging problem…collecting and curating training datasets may not be feasible or practical in many situations, especially as trail conditions may change due to seasonal weather variations, storms, and natural erosion.”
Ground cover such as gravel, dirt, and mulch in addition to seasonal weather can pose serious challenges to a robot. Even dense forests and foliage can wreak havoc on a machine if they can’t adapt.
Researchers saved themselves the tedious process of gathering real-world data to train the robots by sourcing synthetic images of virtual outdoor trails. To create the trails, the team constructed an alpine mountain scene complete with dirt trails in Unity and 3D models of trees, rocks, and grass from the Unity Asset Store. Following the assembly, they released a virtual robot with three cameras containing a 400 x 400 pixel resolution and an 80-degree field of view, that collected 20,269 images of the landscape.
By resizing the images to 100 x 100 x 3 pixels, the team was able to achieve faster processing with lower memory consumption. Following that step, they organized the collection into three sets, one for training, one for validation, and a third for testing. The dataset was fed to three different neural networks. The team utilized, “a deep neural network (DNN), a convolutional neural network (CNN), and a recurrent neural network (RNN).”
During the testing when the AI had control over the virtual robot, the RNN model was able to predict the correct trail direction with 95.02 percent accuracy. “[W]e observed that the robot was largely successful in navigating trails, including those with tight turns and obstacles such as large boulders,” the researchers state in their paper. “Moreover, we observed several instances of ‘intelligent’ decision-making; in one trial, the robot briefly navigates off the trail after colliding with a large obstruction, but then navigates back to the trail and resumes its travel.”
Even with all the progress, none of the systems had a 100 percent accuracy rate. Robots misidentified terrains as trail features, causing them to go off course. While their research isn’t perfect, the researchers believe they have laid the foundation for methods that could enhance AI training in the robots.