Oct 10, 2019 15:00:00

Google announces a benchmark robot that measures the ability of AI for cheap robots

The ability of AI systems installed in robots is measured by benchmark robots, but existing benchmark robots are designed for `` industrial robots that operate with high performance and human monitoring '' Was almost. Therefore, Google

Brain , Google's artificial intelligence research facility, has newly developed a benchmark robot for AI systems that are installed in inexpensive learning-based robots. In order to accelerate the development of the field of reinforcement learning, Google has released a benchmark robot called “ROBEL” in open source software.

[1909.11639] ROBEL: Robotics Benchmarks for Learning with Low-Cost Robots
https://arxiv.org/abs/1909.11639

Google AI Blog: ROBEL: Robotics Benchmarks for Learning with Low-Cost Robots
https://ai.googleblog.com/2019/10/robel-robotics-benchmarks-for-learning.html

In fact, many benchmark tests that measure AI system performance as of 2019 are limited to high-cost, industrial-quality robots. Such robots are evaluated in terms of precision, reproducibility, and ease of use under human control in a controlled environment, but on the other hand, 'recovery for learning in the real world' Many learning-based robots have also been developed in recent years, which have been evaluated for the fact that they are easy to maintain and “reliable reset mechanisms to reduce human monitoring”.

Therefore, Google Brain, Google's artificial intelligence research facility, has newly developed a benchmark robot called “ROBEL” for an AI system installed in an inexpensive learning-based robot. In order to accelerate the development of the field of reinforcement learning, Google publishes ROBEL in open source.

ROBEL
https://sites.google.com/view/roboticsbenchmarks/

ROBEL has two three-arm D'Claw and four-arm D'Kitty made by Trossen Robotics , priced at $ 3200 (about 340,000 yen) and $ 3700 (about 400,000 yen), respectively. D'Claw is 9 degrees of freedom (9DoF) and D'Kitty is 12 degrees of freedom (12DoF). According to researchers, inexpensive and modular ROBELs are useful for performing reinforcement learning on hardware from scratch.

ROBEL's benchmark is D'Claw's task like 'screwing' ...

D'Kitty measures tasks such as 'stand and walk'. Robots often perform tasks such as tightening valves in factories, and robots that are used in places with poor scaffolding require the level of quadrupedal walking.

For example, industrial robots Baxter and Sawyer are highly regarded in the laboratory environment, but the price is as high as $ 15,000 (about 1.6 million yen). However, FRANKA EMIKA, a robot of about 10,000 dollars (about 1 million yen), has appeared, and Resync Robotics , the developer of Baxter and Sawyer, has been acquired by the Hahn Group at the end of 2018. The last few months have been a big challenge, especially for robot engineers, to reduce costs and achieve affordable prices, and researchers at the University of California, Berkeley will have $ 5,000 (2 Announced a robot of less than (10,000 yen) and received a lot of attention.