Khosla-backed robotics startup Genesis AI has gone full-stack, demo shows - TechCrunch
“Our idea was that if we could design a robotic hand that tries to mimic a human hand as much as possible, we can instantly unlock huge amounts of human data without having to worry about what people call the ‘embodiment gap’ in robotics research,” Xian said.
Key takeaways
The most recent headlines show a surge of activity around humanoid robots. In early May, Figure AI unveiled a video of two of its Helix 02 humanoids making a bed together, highlighting new training that lets the robots open doors, push furniture and drape clothing, though the company has not announced a consumer launch date. Figure AI, valued at $39 billion, is racing against rivals such as Tesla’s Optimus and other emerging makers. At the same time, Genesis AI introduced its GENE‑26.5 “brain,” a foundation model paired with a human‑hand‑shaped robotic gripper that it says overcomes the data bottleneck that has limited dexterous manipulation, allowing robots to perform tasks previously possible only with human hands. The company demonstrated the system on both Intel and NVIDIA hardware and is building a massive “human skill library” from egocentric video and internet footage to train the model. In China, Unitree released the GD01, a ten‑foot‑tall, half‑ton mech that can switch between bipedal and quadrupedal modes and is priced at $650 000, marking a bold entry into personal‑size humanoid platforms. Meanwhile, a Tokyo university opened an unmanned laboratory featuring ten robots, including the Maholo LabDroid humanoid, to automate medical experiments and eventually conduct entire research projects without human staff. These developments illustrate a rapid expansion of humanoid capabilities, from household assistance and industrial dexterity to large‑scale, autonomous research environments.
“Our idea was that if we could design a robotic hand that tries to mimic a human hand as much as possible, we can instantly unlock huge amounts of human data without having to worry about what people call the ‘embodiment gap’ in robotics research,” Xian said.
Others have tried their hand at that problem; the main novelty is how Genesis combines this with its model. The current version is named GENE-26.5 for May 2026, but Xian expects there will be many iterations, thanks to the simulation it has developed. “The real bottleneck for the iteration speed of the model is evaluation. So this helps us speed up model training a lot,” he said. Still, it remains to be seen whether workers would be happy to wear the very gloves and cameras that could train robots to replace them, and whether they will get extra pay for that training. That will be between Genesis’ customers and their employees, Gervet suggested. “We haven’t nailed the details yet,” he said.
Either way, they may decide not to share that data with the startup, the founders acknowledged. But the startup also has avenues of its own to build its ‘human skill library’ — it could also pay third-party partners to collect data. Its model is already trained on “massive amounts of human-based internet videos,” according to a press release that didn’t mention compensation. Other well-funded companies operate at the intersection of AI and robotics — such as Physical Intelligence and Skild AI. Zhian also acknowledged that “there’s probably 50 or 100 robotic hand companies out there.” But he and his cofounder Théophile Gervet hope that building their own will give them the upper hand.
The key difference is that Genesis’ hand has the same size and shape as a human hand — rather than the two-finger grippers many robotics companies have been using — reducing the gap with real-world conditions.
“That lets us collect a lot more data than was previously possible, to train a model that can do many more tasks,” said Gervet, a former research scientist at Mistral AI who is now Genesis’ president.