DD-PPO (an AI system for navigation of robots)

2020: Disclosure of source codes

In January, 2020 Facebook opened source codes of the DD-PPO system (Decentralized Distributed Proximal Policy Optimization) allowing be guided robots on premises without use of specially marked cards. According to developers, their technology will open the road for creation of smarter robots and voice assistants.

Systems with artificial intelligence, such as industrial robots, require the detailed map of the area according to which they move to avoid the wrong movements. The problem is that in many cases it is impossible to receive ideally exact design of the premises, people can change location of objects there.

Facebook opened source codes of an AI system for movement of robots on premises

The majority of real environments do not remain constants — buildings and structures change, objects move, and people and pets are in the permanent movement — researchers of Facebook Abhishek Kadian and Erik Wijmans explain.

Facebook developed the algorithms allocating robots with capability independently to be guided in space by means of the built-in sensors. Own testing of the company showed that the program allows a system to reach precisely the destination 99.9% of cases, using only the camera, a compass and GPS. The former record made 92%.

The technology of Facebook trained for 2.5 billion cycles (it approximately corresponds to 80 years of human experience), has important practical value: for the real robot acting in conditions of office or laboratory even the unique wrong movement from hundred bears risk of damage of the equipment and therefore it is inadmissible.

Besides that the number of the wrong turns is reduced almost to zero, "our agent also shows capability to make reasonable decisions, such as choice of the correct fork on the road", developers stated.^[1]

Robotics

Notes