A team of engineers at Google’s DeepMind Project has demonstrated a robot capable of playing amateur-level table tennis (ping-pong). The team has published a paper on the arXiv preprint server describing how they developed the robot, how well it performed at different ability levels and how human players responded to playing with the robot.

Over the past several years, robot scientists have been combining advancements in robot design with artificial intelligence, resulting in the development of robots with ever increasing abilities. In this new effort, the research team has developed an AI-based ping-pong player with the highest performance level ever for a robot.

To build their robot system, the researchers started with a robot arm called the ABB IRB 1100—the robot is currently used in real-world industrial applications. In addition to its ability to manipulate its arm and hand very quickly, it can also quickly slide side-to-side on a rail. These features made it an ideal ping-pong-playing candidate.

Picture this: hundreds of ant-sized robots climb over rubble, under rocks and between debris to inspect the damage of a fallen building before human rescuers explore on-site.

Downscaling legged robots to the size of an insect enables access to small spaces that humans and large robots cannot reach. A swarm of small robots can even collaborate like their insect counterparts to haul objects and protect one another. Picotaur, a new robot from the labs of Sarah Bergbreiter and Aaron Johnson is the first of its size, able to run, turn, push loads and climb miniature stairs.

“This robot has legs that are driven by multiple actuators so it can achieve various locomotion capabilities,” said Sukjun Kim, a recent Ph.D. graduate advised by Bergbreiter. “With multiple gait patterns, it can walk like other hexapod robots, similar to how a cockroach moves, but it can also hop from the ground to overcome obstacles.”

A trio of chemists at the University of Copenhagen has developed an AI application that can be used to figure out the phase of x-rays that crystals have diffracted as part of efforts to predict the structure of small molecules.

To address challenges of training spiking neural networks (SNNs) at scale, the authors propose a scalable, approximation-free training method for deep SNNs using time-to-first-spike coding. They demonstrate enhanced performance and energy efficiency for neuromorphic hardware.