Genesis AI Claims 43 Million Frames per Second; Independent Test Finds 0.29 Million

Genesis AI unveiled a robot hand this week that can crack eggs, solve Rubik's Cubes, and play piano at 130 beats per minute. The company is also marketing something else: a speed claim it says makes its entire approach viable. Genesis says its physics simulation engine runs at 43 million frames per second, roughly 430,000 times faster than real time. An independent researcher ran the actual benchmark and got a different number: 0.29 million frames per second. That is 150 times slower than Genesis advertised.

The finding, published by researcher Stone Tao on his blog in December and still the most detailed independent analysis available, concluded that Genesis's headline figure was achieved by benchmarking a robot arm that sat idle more than 90 percent of the time, with physics accuracy settings turned down and self-collision detection disabled. When Tao applied the same benchmark to a genuine robot manipulation task, Genesis ran three to ten times slower than ManiSkill, an open-source PhysX-based simulator that robotics researchers already use. A Genesis author later confirmed on social media that when you run the same locomotion training task in Isaac Lab, it trains at the same speed as Genesis. There is no speed advantage. The company disputes Tao's methodology and says its latest simulation version achieves higher throughput, but has not published third-party verified benchmarks.

Genesis AI raised $105 million in seed funding last July from Eclipse and Khosla Ventures, with Eric Schmidt and Xavier Niel as notable individual backers, according to PR Newswire. The company says it is in advanced talks with potential customers in France, Germany, and Italy, targeting sectors like automotive, electronics, and pharmaceuticals where conventional robots struggle with delicate or variable tasks, per Reuters. The pitch is that its data engine, combining real-world robot interaction, high-fidelity simulation, and internet-scale video data, solves the data bottleneck that has kept more than 95 percent of physical labor unautomated.

The robot hand itself is technically interesting. It has 20 degrees of freedom and 20 motors housed directly inside the hand, rather than using the tendon-driven approach common in industrial arms. CEO Zhou Xian, a Carnegie Mellon PhD in robotics, told Business Insider the hand moves at 60 to 70 percent of human speed. In cooking demonstrations, most tasks achieved 90 to 95 percent success rates, but one-handed egg cracking and transferring chopped tomato with a knife fell to 50 to 60 percent during filming. The demos were real, the speeds are real, and the hand design is a genuine departure from the norm.

The question is whether the simulation stack that trains these hands is actually the competitive moat Genesis claims. The underlying physics engine, Genesis, was released as an open-source academic project by researchers from Carnegie Mellon, MIT, Stanford, Columbia, and UMD before the company commercialized it, per the Genesis research platform. It is not a new product. The open-source version is available to anyone. What Genesis added was the commercial wrapper and the claim of superior speed. If that speed claim is inflated, the data engine advantage and the customer pitch built on it are correspondingly weaker.

Physical Intelligence, Figure AI, and 1X are all building robot foundation models and data pipelines. None are publicly claiming 430,000-times-real-time simulation throughput. The robotics industry has long struggled with the gap between simulation and reality — training in a virtual world does not guarantee a robot performs the same task at the same speed outside it. Genesis says its rendering engine and hyper-realistic physics bridge that gap. The independent benchmark suggests the gap remains.

Genesis says it will share more data on its simulation performance and has published an updated benchmarking report since Tao's critique. It declined to provide third-party verified results for this article. The robot demos are real. The hand is real. The speed claim that underpins the rest of the story is the part worth checking.