Get best of the week

A new era in robotics has begun



CoSTAR team with four-legged robot SPOT mini won the Urban Circuit stage of the DARPA Subterranean Challenge!

Robot competitions evolve


The Department of Defense Advanced Research Projects Agency ( DARPA ) is the agency responsible for developing new technologies for military use. According to a quote from the DARPA website, “to ensure the continued technological superiority of the US armed forces, to prevent the sudden emergence of new technical means of warfare for the United States, to support breakthrough research, to bridge the gap between basic research and their application in the military sphere, the agency conducts a number of events in including competition among robotic companies. ” The DARPA Subterranean Challenge is one such event.

Unmanned competition “Winter city”showed that the teams used the trick, used to localize on the GPS track and did not go to those parts of the track where its signal is unavailable, thereby invalidating the idea of ​​revealing the level of readiness of UAV technology. In the DARPA Subterranean Challenge, the action takes place underground, excluding the very possibility of using GPS, and there are other obstacles that cause problems for robots and AI: poor lighting, unstructured environment, puddles, stairs between floors, etc. The task is for the robots to autonomously explore the dungeon and find certain objects, for which they are awarded points. Thus, the DARPA Subterranean Challenge forces us to look for new approaches for quick mapping,underground navigation and search during time-sensitive combat operations or disaster response scenarios.



On February 27, the CoSTAR team with the four-legged walking robot SPOT took first place in the Urban Circuit of the DARPA Subterranean Challenge, while the CERBERUS team with its four-legged walking robot ANYmal took fifth, while walking robots were first used in such competitions.

CoSTAR Team Victory and Technology Analysis



1. In order for robots to be truly autonomous in conditions of movement in dungeons without access to a GPS signal, the team developed the NeBula framework using fuse data from various sensors and detecting anomalies in them using ML. The use of new technology will allow robots to perform critical tasks offline in difficult conditions that are now "too tough" for the current generation of robots and unmanned vehicles. It is important that this NeBula framework is implemented in the Robot Operating System (ROS), which is one of the basic technologies in the Sberbank Robotics Laboratory and is actively developed through the developer community. The next meeting of several hundred ROS engineers will be held at Sberbank on April 18 .
2. Overcoming human infrastructure and an unstructured environment, such as flights of stairs, is currently a serious problem for robots. On a section with a staircase between floors, the SPOT robot demonstrated its capabilities in the best way (a short passage where the robot descends the stairs). Passing such a test suggests that it now becomes possible to create more complex autonomous robots for moving simultaneously both on the street in difficult terrain and inside rooms not prepared for robots.

CERBERUS Team Analysis



Video: ANYmal robot descends the stairs
ANYmal including a research platform on which open research is conducted and scientific articles are written, one of the last approaches that was applied in competitions is described in the article “Learning agile and dynamic motor skills for legged robots”.It consists in providing an approach to solving the problem of machine learning in a simulator and in reality. The neural network was trained on the basis of a mathematical model of leg movement and data collected from a real robot, then it was trained on another neural network based on the predictions of the first neural network in the simulator. The hybrid simulator turned out to be faster and more accurate than the simulator on analytical models. But more importantly, when the movement strategy was optimized in a hybrid simulator, and then transferred to a real robot and tested in the physical world, it turned out to be as successful as in the simulation. This long-overdue breakthrough signals a decline in the seemingly insurmountable gap in the training of neural networks between simulation and reality.



The approach used hints at another important shift in the field of robotics. Hybrid models are the first step to this change. The next step will be the rejection of analytical models in general in favor of machine learning models that are trained using data collected in a real robot environment. Such data-based approaches called end-to-end are gaining momentum.

Also, the development of such approaches will help to approach the solution of AGI tasks, following the example of how a child learns to walk. Thus, the study of robotic software can provide insights into long-standing questions about the human mind.

It can be assumed that self-awareness and therefore consciousness, in essence, are an indicator of our ability to think about ourselves abstractly - to express ourselves. The further a person can look forward, and the more detailed the mental picture of his future activity will be, the higher will be the ability of this person to self-consciousness. Now robots are able to learn to independently model. This breakthrough is not only a practical achievement that will save some engineering effort, but also the beginning of the era of robot autonomy.

Authors of the article: Albert Efimov, Alexey Burkov, Victor Tsygankov
Sberbank Robotics Laboratory

More articles:


All Articles