Two people with direct knowledge of the matter said that Ashish Kumar, former AI lead for Tesla's Optimus humanoid robot, has co-founded a startup to develop non-humanoid robots. These robots can take on tasks in manufacturing, aerospace, and other industries that previously required highly skilled workers.

The San Francisco-based startup is called Intelligent Machines and has recruited several former members of Tesla's Optimus team. It is currently raising a seed round that could reach about $100 million.

Intelligent Machines is currently focused on robot software development, but its management is also discussing building its own hardware. Two people with direct knowledge said the company uses reinforcement learning to train robots: rewarding them when they achieve goals and penalizing wrong actions, thereby teaching them to perform complex tasks. (Kumar's doctoral research at UC Berkeley was precisely in reinforcement learning.)

Reinforcement learning has long been used in robot training, but in the past it was mostly limited to teaching robots specific skills, such as walking or playing table tennis. More recently, robot researchers have begun applying reinforcement learning to more practical tasks, such as giving positive rewards for successfully inserting a cable or folding a napkin.

One person with direct knowledge said Kumar believes the latest robot AI models currently cannot provide the precise control needed for physical automation of manual labor in industries such as manufacturing. It is not yet clear whether Intelligent Machines uses vision-language-action models, or VLA models. Such models are derived from the large language models that power AI chatbots and are specially trained to control robots. But it is certain that the company's technical approach differs from that of other startups developing so-called world models.

World models are trained on video data, understand physical laws, and simulate the real-world environments in which physical robots operate. Companies in the world model track include General Intuition, which completed a $220 million financing last month at a valuation of $6.2 billion, and World Labs, which chipmaker AMD plans to acquire for about $8.2 billion.

Although capital continues to flow into robot startups, current robot AI models have yet to see a ChatGPT-like breakout moment, meaning a moment when many kinds of robots can autonomously complete a large variety of physical tasks. Venture capitalist Vinod Khosla said in an interview last month that while he believes this moment will arrive within two years, he also predicted that by 2030 more than half of robot startups will see their valuations shrink.

A person familiar with the startup's R&D work said that although Intelligent Machines' robot prototype is similar to Optimus in some respects, it is not a humanoid robot, and in terms of AI logic it has a "major difference" from Optimus.

Tesla has encountered engineering difficulties in developing Optimus's robotic hand, causing production delays and hindering efforts to scale up robot output. Kumar worked at Tesla for more than two years, including one year as AI lead for the Optimus team, and left a year ago. (His LinkedIn profile shows that after leaving Tesla, he served as a research scientist at Meta for four months.) Kumar's former colleague on the Optimus team, Kamal Gupta, also announced on LinkedIn this summer that he had left Tesla.

Intelligent Machines is the latest in a series of startups founded by former Tesla employees. Last month it was reported that AI chip company DensityAI, founded by former core members of Tesla's Dojo supercomputer project, is in late-stage financing talks to raise several hundred million dollars at a post-money valuation that could reach $10 billion.