Movie Reviews

This Robot Dog Can Chase You Off-Road Thanks To Computer Vision

Many people will remember Boston Dynamics’ robot named Spot, but a new four-legged robot dog being developed by researchers at Carnegie Mellon University and UC Berkeley is now in the spotlight. Robotics and technology have come a long way over the last few years, and researchers are constantly working to create better and more sophisticated robots that are powered by artificial intelligence.

Unlike most animals that tend to rely on their vision to move about their surroundings, robots can use a combination of sensors and cameras to map out their environment. Boston Dynamics’ robot for instance uses an internal map to move about. However, this can cause some constraints, such as robots not being able to navigate new surroundings and obstacles. Researchers want to change that with a new four-legged robot dog.

Related: Xiaomi Beats Tesla To A Bipedal Robot Prototype

The team at Carnegie Mellon (via MIT Technology Review) have created a robot dog that relies exclusively on cameras to guide its movements. Before being released into a new environment, the robot is first trained to navigate different environments in a simulator, much like how human babies learn to walk using trial and error. Since the robot only uses vision to navigate the terrain, it must remember the location of objects behind it while moving. A short video shows the robot climbing over barstools, where it needs to remember the location of the stools it just stepped over to place its hind feet correctly.

The researchers said the robot also demonstrated “emergent climbing up behavior,” which allowed it to solve problems even after initially getting stuck. Since the robot relies on vision, it can move across terrain that it has not been specifically trained for in a generalized manner, similar to how humans move across new terrain they have never seen before. The robot is extremely resilient as well, being able to walk across slippery surfaces and survive falls.

Of course, nothing is perfect, and there are some limitations and failure cases for this robot. Firstly, since the robot uses human-like vision to move, that comes with human-prone mistakes as well. For example, when moving across multiple objects with gaps between them, the robot is not able to see the objects behind it. If there is an error in retrieving previous positions of objects that the back legs need to step on, the robot will miss the step and fall over. Another problem happens with a step that is too large and when the extent of the drop cannot be seen. When this happens, the robot simply falls off the step. But with more refinement and corrections, these problems can be solved in the future.

Interestingly though, when the robot was placed in completely unstructured terrain that it hasn’t been trained in, such as rocky or slippery slopes, it was still able to adapt and move forward, even though it had accidents along the way. The robot’s camera also includes an infrared light sensor that allows it to navigate in the dark with only a little ambient light. Videos show it crossing roads and climbing down stairs at night, which is pretty impressive. While it still has a long way to go, the team at Carnegie Mellon have taken a major step towards creating a robot that can survive in the real world.

Next: Uncanny Valley Theory: Robots With Human Characteristics Being Embraced

Source: MIT Technology Review, VisionLocomotion

Back to top button