The cornhole run, throw by throw
Google said its Apollo humanoid sank seven straight cornhole shots. The engineer who ran it then posted the whole log: throw #1 landed 85 cm short because the grip was slipping, and the fixes came one throw at a time.
The big one
Physical Intelligence opens its Pi API, from strawberries to servers
Physical Intelligence put the Pi API into beta, saying it has already partnered with companies to adapt its models to their robots, from harvesting strawberries to assembling servers. Chelsea Finn says many people have already been fine-tuning Pi models for their own robots and the team wants to expand testing significantly. Lerrel Pinto says the striking part is how many different types of robots now run the same intelligence layer.
quite exciting to see so many different types of robots using the same intelligence layer
Shown this week
The Gemini cornhole run, throw by throw, from the engineer who ran it
Google said Gemini 4 Argon watched a video of a human playing cornhole, autonomously built its own simulation code to practice, and then Apptronik's Apollo sank seven straight shots.
Large models really are changing a lot of the assumptions about how intelligence for robotics is going to work
AlphaNIST solves all 20 NIST assembly tasks at 98.2% in simulation
Octi Zhang released AlphaNIST, a single policy trained from scratch on a binary success or failure reward that he says solves all 20 NIST assembly tasks at 98.2% success in simulation. It can also run 10 different tasks back to back and only get the reward once, at the end.
First promising result for sim2real in manipulation
Out in the world
Origin Autonomy says its drywall robot already runs on New York sites
Origin Autonomy came out of stealth with $20M and OG-1, a robot that Lukas Ziegler says finishes drywall autonomously at three times human speed. Per his post, the first robots went onto New York job sites within 10 months of founding, and the target is 150 deployed robots by the end of 2027. OG-1 applies compound, sands, primes and handles corners, pillars and soffits at heights up to 14 feet.
Teleop and other arguments
30,000 hours of video taught a world model hands, not objects
Jiahua Dong says his team trained a world model on 30,000 hours of first-person video and it learned exactly where hands go but not what happens to the things they touch, and going from 300 to 30,000 hours did not close the gap. A 4x larger model barely helped either.
Sorry man, perception is deception.
Frequently asked questions
What is the Pi API and what can it do?
The Pi API is Physical Intelligence's model interface now in beta that can be adapted to different types of robots for tasks like strawberry harvesting and server assembly. Multiple companies are already fine-tuning Pi models for their own robots.
How did Google's Gemini get a robot to make cornhole shots?
Gemini 4 Argon watched a video of cornhole, built its own simulation code to practice, and then Apptronik's Apollo robot sank seven shots in a row. The robot used classical simulation tools like MuJoCo to improve through repeated practice.
What is NIST assembly and what does AlphaNIST do?
NIST assembly tasks are a set of twenty standardized robotic assembly challenges. AlphaNIST is a single policy that solved all 20 tasks at 98.2% success rate in simulation and can run 10 different tasks back to back.
How fast is Origin's drywall robot compared to humans?
Origin Autonomy's OG-1 robot finishes drywall three times faster than human workers and can apply compound, sand, prime, and handle corners, pillars and soffits at heights up to 14 feet. The company deployed its first robots to New York job sites within 10 months of founding.
Why is training on video not enough to teach robots to manipulate objects?
A world model trained on 30,000 hours of first-person video learned where hands go but not what happens to objects they touch, and increasing training data from 300 to 30,000 hours did not close that gap. Even a 4x larger model barely improved the results.