Did you ever imagine whether it be a mobile robot and autonomous systems moving safely on sidewalks to avoid collisions with obstacles or pedestrians? In order to operate effectively in the urban space, few of the delivery robots or systems are being programmed to patrol urban environments.
For this very reason AlienGo, a quadruped robot was developed by researchers at the Georgia Institute of technology and Stanford University recently. AlienGo can follow specific routes that are generated by public map services by staying on the roadside without getting collided with humans or obstacles.
AlienGo was presented in a paper that was pre-published on arXiv, which is based on a new, highly performing two-staged learning framework that is used for safe sidewalk navigation. One of the researchers from the team, Sehoon Ha says that as part of this project they have developed and designed an intelligent quadrupedal robot which is capable of navigating sidewalks in the real-world setting. He further says that their work has been inspired by two stems of the existing work one being autonomous driving and other being the indoor robot navigation.
However, outdoor sidewalk navigation typically takes place in an unstructured space with a number of pedestrians and obstacles without any guide lanes. Sehoon says that they have also proposed a set of learning techniques and algorithms to solve a few of the specific problems.
At the initial stages, the team of researchers has trained an artificial neural network to navigate simple sidewalks and ecosystems in simulations. While coming to the first algorithm, it was dubbed as ‘expert’ and was given training using a high-speed salient world simulator with access to the so-called ‘privileged state’ of the simulation.
When this expert network was transferred to a student algorithm in a high-fidelity simulation, it produced realistic sensor observations that mimicked real-world sidewalk images. Another researcher of the team, Maks Sorokin, said that the student uses a custom-trained semantic feature network to generate all the abstractions which are used to control the robot. He adds saying that this experience was based on their experience that the desired behavior is difficult to obtain using naive end-to-end training as the problem is a bit hard to obtain.
Using the two-stage learning framework the researchers have developed, they were able to attain an effective policy using privileged information in simulation and then transfer the behaviors acquired by the framework to a real four-legged robot. The researchers also tested the AlienGo in the real-life setting as it navigated sidewalks in Atlanta.
The quadrupedal robot which was developed by the team of researchers can deliver parcels, monitor urban environments, and also can be useful in various tasks. They designed and developed mobile robots to improve their ability to navigate sidewalks.
Sorokin says that although they have made progress in sim-to-real transfer for navigation, there are still many challenges remaining. “Some of the challenges related to navigation that we still need to overcome include road crossing, dynamic obstacle handling, and interaction with real-world objects and humans, However, our approach is not limited to navigation, it could potentially be applied in many robotic applications, such as manipulation, locomotion, and others. We are excited to see its applications in adjacent research areas.”