Inside the Montreal lab training robots to work on real factory floors

Key points

• Vention opened its Physical AI Lab in Montreal in 2026, focusing on teaching robots to handle complex manufacturing tasks.
• The company has deployed more than 28,000 machines across a community of more than 6,000 factories worldwide, including 90 of the Fortune 500.
• Vention's physical AI revenue grew 400% over the past year, according to the company.
• Vention launched GRIIP (Generalized Robotic Industrial Intelligence Pipeline) in February 2026, a modular software system for guiding robots through picking, gripping and assembling objects.
• Dr. Joelle Pineau, chief AI officer at Cohere and former VP of research at Meta, joined as an external technical advisor.

A robot that works perfectly in a lab can still fail spectacularly on a real factory floor. Noise, variation, unexpected parts: the real world is messier than any controlled test. Vention, a Montreal-based automation company, thinks the answer is to skip the lab-to-factory gap entirely and do the research inside the factory itself.

The company opened its Physical AI Lab in Montreal this year, as first reported by The Robot Report. The goal is to train AI models that control physical robots using data collected from actual production lines, not synthetic simulations.

What does the lab actually do?

The lab feeds real industrial data directly into AI model training, then sends improved robot behaviour straight back to factory floors. Vention's CEO Etienne Lacroix put it plainly: "Any physical AI foundation model is data-hungry." With more than 28,000 machines already running in client factories, the company collects a continuous stream of real-world robot movement data that most AI research teams simply cannot access.

The lab works on tasks that sound simple but are genuinely hard for machines. "Kitting" is one example: taking a jumble of components from different suppliers and assembling them into a specific set that a worker needs at a particular station on the assembly line. A car plant might need headlamps, wiring harnesses, mounting brackets and screws bundled together in exactly the right configuration before the car reaches position seven on the line. Getting a robot to do that reliably, across thousands of variations, is an open research problem.

Dr. Jimmy Li, a robotics researcher from McGill University who leads the lab, described the unusual setup: "Our clients aren't waiting for a finished product to test; they're in the room while we build it."

What is GRIIP, and why does it matter?

GRIIP is Vention's modular pipeline, a chain of software steps, for guiding a robot arm through the full sequence of seeing an object, identifying it, calculating how to grip it, and moving without hitting anything. Launched in February 2026, it combines AI foundation models, large general-purpose models trained on vast amounts of data, from partners including NVIDIA with Vention's own proprietary models. Vention plans to release an open-source version of GRIIP so other developers can build on it.

The company is also working with a large automotive manufacturer on high-complexity assembly tasks, and with electronics contract manufacturers who handle components from dozens of different vendors, all arriving in different packaging.

What does this mean for manufacturers?

For factory owners, the pitch is simpler and cheaper robots that can handle more variety without needing a specialist to reprogram them every time something changes. Vention expects the cost of deploying AI-driven robots to fall as the underlying models mature, opening the technology to smaller manufacturers that could not previously justify the investment.