Ping Pong Robots, Weird AI businesses, & Agent Gating

TLDR;

Welcome back! This week, we have another exciting week, with a ping pong robot coming from China that will leave you speechless, the latest episode in the OpenAI vs xAI drama, Cohere’s wise strategy, and CloudFlare’s agent gating.

Moving on, we’ll analyze the fascinating business model of the latest success story in AI: OpenEvidence, among other news coming from OpenAI, Google, or Microsoft.

Enjoy!

ROBOTICS

Ping Pong Robots

In one of the most fascinating announcements in a while, a group of Chinese researchers has presented HITTER. This new method allows robots to play up to 100 consecutive shots at ping pong, an incredible milestone for the AI robotics industry.

But how does it work?

HITTER is composed of two parts:

A model-based planner first predicts the ball’s trajectory using physics equations and determines when and how the racket should strike it, specifying targets like strike time, racket pose, and velocity. These high-level plans act as guidance, ensuring the robot always has a physically accurate strategy for returning the ball.

Execution is then handled by a learning-based whole-body controller, a neural network policy trained via reinforcement learning with human demonstrations as priors. The policy learns to map the planner’s strike targets, along with the robot’s body state, into joint-level actions that produce natural, feasible, and stable motions.

In other words, a physics engine (this is a deterministic symbolic engine, no AI here) computes the ball’s trajectory. The robot’s AI policy, in charge of moving it, takes in the planner’s ball data plus the robot’s own sensory data (racket and body positions) and, through rewards shaped around achieving reference poses, velocities, and correct timing, it enables the robot to translate abstract strike strategies into precise whole-body movements that consistently return the ball.

→ WATCH VIDEO

TheWhiteBox’s takeaway:

The more someone is familiarized with the complexities of robotics and AI, the more impressed they are with this news. Almost a miracle they work.

Let’s recall that the Moravec paradox governs AI robotics; what appears easy for you is hard for humans, and vice versa. For us, bodily movements are easy, natural. But that is the exact opposite for machines, which have to deal with dozens of sensory inputs and multiple-hundred Hertz frequencies (the robots need to output hundreds of movements per second). This is the main reason robotics AI is an extremely challenging field.

In my humble opinion, the reason the Moravec paradox exists in the first place is because of the extreme data scarcity in robotics. No data, no AI.

But the biggest takeaway is that these researchers are taking an “unconventional” approach here. While most US Labs playing the robotics game aim for an end-to-end AI solution, these researchers took the smart route and avoided using AI in the hardest area: ball trajectory prediction.

Instead, OpenAI researchers would tell you that the aim is for the AI to do everything, from predicting ball movements to meeting these with the robot’s racket.

More clearly described, what the Chinese team is doing…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.