Vention Launches Physical AI Lab for Factory Robots

Vention, a Montreal-headquartered robotics firm, has set up a new research lab focused on physical AI that will use data generated by its machines on real factory floors to improve what robots can do and how well they do it, according to The Logic's reporting on the launch.

Vention's strategic positioning here is genuinely distinctive. The company has no plans to compete against startups that have raised huge sums to develop multi-purpose foundation models for robotics. Instead, Vention is focusing specifically on improving software and control systems based on how robots are actually used in real industrial settings, according to CEO Etienne Lacroix. "It's not profitable to deploy physical AI today," Lacroix said. "We can deploy it at scale with amazing unit economics."

The company's core differentiation rests on the source of its training data specifically. Vention will draw on data gathered by tens of thousands of machines it has sold to manufacturers across North America and Europe since starting in 2016. Many physical AI startups train their foundation or world models using scraped videos, or by paying workers to generate examples wearing sensor-covered gloves or point-of-view cameras. Lacroix claims Vention's data is more effective specifically because it comes directly from robots already operating on real production and assembly lines, not simulated or human-generated proxies for that same work.

Vention's Physical AI Lab at a Glance

One genuinely concrete challenge the lab is tackling involves training robots to pull items from deep bins without colliding with nearby machinery or workers, a problem that has historically restricted many industrial robots to slow, simple conveyor belts or lifting and dropping items within closed safety cages. "Our clients are just more demanding, the use case is a bit more tricky," Lacroix said. The team is also working to automate kitting, the process of unboxing and gathering different pieces for a specific task, such as a headlamp, harnesses, brackets, and screws on an automotive assembly line, technology Vention has already tested with a major automaker.

Vention's software, called GRIIP, runs on a mix of commercial and open-source AI models combined with proprietary algorithms developed in-house, designed to run on the controller unit Vention already sells with its robots, without requiring any hardware upgrades. The company charges clients $15,000 USD annually per machine unit to use GRIIP, a business model Lacroix argues beats trying to develop a foundation model outright, which he says can command only a few thousand dollars of revenue per robot. "We really see those base models as a commodity," he said. "There's so many people right now developing them."

Vention raised a $110 million USD Series D round in January at a valuation exceeding $1 billion, backed by Investissement Québec, Desjardins Capital, and Nvidia's venture arm, according to The Logic's earlier reporting on that raise. The company operates in a genuinely crowded field of well-funded physical AI competitors taking the foundation model approach Vention has explicitly decided to avoid, including Pittsburgh-based Skild AI, which closed a $1.4 billion round from SoftBank and Nvidia, San Francisco-based Physical Intelligence, reportedly valued at $5.6 billion, and Montreal's own Veeda AI, which we covered in detail in our earlier reporting on former Nvidia researcher Sanja Fidler raising over $90 million for a robotics world model startup.

Vention's approach is worth understanding for any business in manufacturing or industrial automation evaluating physical AI vendors, since it represents a genuinely different strategic bet than the foundation-model approach dominating most robotics AI headlines. A company with real, deployed data from tens of thousands of production machines and existing enterprise customers like Boeing and 3M offers a fundamentally different value proposition than a startup still building general-purpose robotics intelligence from scratch.

For businesses evaluating whether to build or buy AI capability for physical operations specifically, Vention's explicit framing of foundation models as a "commodity" rather than a defensible business is worth considering, since it suggests the real, durable value in industrial AI may increasingly concentrate in applied, use-case-specific software rather than the underlying general-purpose models themselves.

What is Vention's new physical AI lab focused on?
Vention's lab uses data from tens of thousands of robots it has already sold to manufacturers to improve robot software and control systems for real industrial tasks, rather than building a general-purpose foundation model.

Why doesn't Vention build its own robotics foundation model?
CEO Etienne Lacroix said Vention views foundation models as a commodity given how many companies are developing them, with limited per-robot revenue potential, compared to Vention's software product GRIIP, which generates $15,000 annually per machine unit deployed.

Who leads Vention's new physical AI lab?
The lab is led by Jimmy Li, Vention's director of physical AI, who holds a doctorate from McGill University, with Cohere Chief AI Officer Joelle Pineau serving as an advisor to the unit.

Vention, a Montreal-based robotics firm, launched a physical AI research lab that uses data from tens of thousands of already-deployed machines to improve robot software for real manufacturing tasks, explicitly avoiding the foundation model race pursued by many well-funded competitors. The 16-person lab, led by director Jimmy Li with Cohere's Joelle Pineau advising, is tackling specific problems including collision-safe bin picking and automated kitting for automotive assembly lines. Vention charges clients $15,000 annually per machine unit for its GRIIP software, a business model CEO Etienne Lacroix argues beats competing against well-capitalized foundation model startups like Skild AI and Physical Intelligence, framing general-purpose robotics models as an increasingly crowded commodity market.