Researchers and Robots Swarm Japan’s RoboCup Competition

by Lynette Farinas

Some robots played soccer. Others tackled picking and packing duties in a simulated warehouse. Still others competed to see which could tackle household chores, like setting the table.

100,000 spectators — including 3,000 students and scientists — joined thousands of robots like these last week in Nagoya, Japan, for RoboCup 2017, the world’s largest robotics competition.

The event’s competitions — held over four days — are part of an effort to advance robotic R&D. NVIDIA was a first-time sponsor, but our technology powered teams competing in each of the week’s three main events.

robocup robots
Put me in, coach: RoboCup 2017 robots ready to play today.

A Field of Robotic Dreams

The robot soccer tournament was a fan favorite.

Crowds of spectators cheered on as pint-sized autonomous robots scored goals — and mastered flops.

We sponsored the CIT Brains team from Chiba Institute of Technology, whose Jetson-powered humanoid robot won the Technology Challenge among all the RoboCup Soccer Humanoid leagues for the KidSize class (40-90cm height).

CIT’s team used Jetson on their robot to locate the soccer ball and make plays, including goal-tending and scoring.

“With Jetson TX1, we are able to implement a deep learning-based object detection network to find the soccer ball and goal post using YOLO — you only look once,” said Youta Seki, team leader of CIT Brains.

The goal is to build a team of humanoid robots that can beat the World Cup championship team by 2050.

Finding a Home for Robotics

For the first time, the RoboCup@Home challenge used the Jetson-based Toyota Human Support Robot (HSR) as one of its standard platforms.

Competitors from 15 universities built features that could assist people in everyday domestic tasks, such as setting the table, hanging clothes and bringing food.

Justin Hart, team lead for the University of Texas at Austin, stressed the importance of accessing the NVIDIA Jetson in the Toyota HSR.

“There’s a lot of systems running inside the robot in parallel,” Hart said. “We’re also able to reduce latency required to process frames of video (that the robot sees) with CUDA.”

His team placed third in the competition, behind Hibikino Machina and Team eRasure, both from Japan.

GPUs and Deep Learning Come Out on Top at Amazon Robotics Challenge

This year, the Amazon Robotics Challenge was held in conjunction with RoboCup 2017.

Since its start three years ago, the competition has grown in importance as Amazon looks to further automate the packaging and shipping of millions of customer orders.

Sixteen universities from top robotics programs competed in this year’s finals, vying for more than $250,000 in cash prizes, as well as bragging rights.

robocup 2017 cartman
On their way to $80K: A broken robot wrist had the team from the Australian Center for Robotics Vision scrambling for a fix in the Amazon Robotics Challenge.

The event challenges autonomous robotic arms to figure out how to pick and stow objects. This year, organizers introduced new items — just 30 minutes before each team’s challenge began — for robots to pick and place into separate Amazon order boxes. With real-time deep learning training and NVIDIA GPUs, one of the teams was able to complete this task in as few as eight minutes.

Cartman, a robot from the Australian Center for Robotics Vision, was one of the Cinderella stories at the event, with laser printed components.

A broken wrist in a practice run had the team scrambling for a fix before the stowing phase. With ingenuity, lots of zip ties and a 3D printer working overtime, Cartman was up and running again, and ultimately took the pencil holder winner’s award, along with $80,000 in cash prizes.

Other winners included MIT, which took the top prize for the stow challenge, and Nanyang Polytechnic University, which won the picking challenge.

All the competing teams used NVIDIA GPUs to train neural network models with thousands of store item images. GPU deep learning inference was also used for object and pose recognition during the competition.