Manipulation is considered one of the hardest problems in robotics. The ability to grab, move, take apart or put together an object with precision is what would make humanoid robots truly useful in the real world.
Loading audio narration…
A startup backed by VC firm Eclipse and former Google CEO Eric Schmidt said it took a big leap to solve it.
Genesis AI, a French startup with an R&D center in Silicon Valley, said on Wednesday that it’s getting closer to achieving “human-level capability” in manipulation, showing a robot playing the piano, cracking an egg, and harnessing wires in recorded demonstrations.
One video showed robotic hands keeping up with a piano composition that moves at around a brisk 130 beats per minute. The startup also demonstrated robots cracking an egg with one hand and harnessing wires.
Genesis AI said the demos were executed autonomously — meaning the robots were not teleoperated by a human — and shown at 1x speed.
The demos were not examples of zero-shot execution, meaning the robots still required training for specific tasks, such as playing a particular piece of music. Zhou Xian, CEO of Genesis AI, told Business Insider in an interview that his team of around 60 people taught the robot to play a new song on the piano in one hour.
For the cooking demonstration, Xian said it required a “few hundred trajectories” or recorded examples of relevant tasks that would train the robot to crack an egg or chop a tomato.
He said a 30-second “complex skill” — such as those seen in the cooking demo — requires a few hours of human data, combined with less than half an hour of data from the robot performing the task.
The robot still failed on some subtasks of the cooking demo: While most steps reached roughly 90% to 95% success, Xian said one-handed egg cracking and transferring chopped tomato with a knife were closer to 50% to 60% during filming.
“I think these are probably the most complex tasks ever being performed by a robot in a very human-like way at the efficiency, speed, and performance similar to a human,” Xian said, adding that the robot is exhibiting about 60% to 70% of human speed.
Genesis aims to build a general-purpose robot capable of performing a range of tasks across different environments. Unlike model-focused firms such as Physical Intelligence, the startup is developing the entire stack: the AI model, robot hand, training gloves, simulator, and, eventually, the robot itself.
In 10 years, Zhou said, he does not see why a factory robot should be fundamentally different from a home robot.
“I think the beauty of being a full-stack company is when you design the hardware, you know exactly what’s needed,” Xian said.
Genesis is making a hand that closely resembles the human form. Xian said the startup’s robot hand has 20 degrees of freedom and 20 motors directly inside it. That differs from tendon-driven hands, where motors are placed in the forearm and cables, or tendons, move the fingers.
Rather than relying solely on video data and teleoperation, the startup is using a mix of internet data and raw human data collected through proprietary training gloves that capture hand motion and tactile, force-like signals.
Xian said his company is talking with a few industrial partners that could have employees wear the training gloves for data collection while at work. Genesis also uses an in-house simulator to test models trained on real-world data across many virtual environments, which Zhou said lets the company evaluate systems faster than running each test on a physical robot.
The CEO said he’s not making the bold claim that manipulation has been solved, but that Genesis’ approach is a “critical step” toward pushing robot manipulation to the next level.
“We’re indeed an ambitious company,” Xian said, “and we’re just not happy with the status quo, and we want to push the field forward.”
