Vention announced a Physical AI Lab in Montreal on September 9, adding a dedicated research facility for teaching factory robots to handle more varied work. Led by Jimmy Li, its director of physical AI, the lab will combine industrial data collection and robot-control research with further training of existing AI models for manufacturing tasks.[1]
The company describes a research programme shaped by production-line requirements, including reliability, cost and variation between parts. It says customers are involved during development. The opening establishes a place to pursue that work, not a demonstrated improvement in factory performance.[1]
Li's accompanying company interview draws an important boundary around today's systems. Vention currently deploys cells using a modular pipeline: separate software stages interpret a scene and plan how the robot should move. He says this approach is becoming easier to set up, but remains difficult to program for dexterous actions such as removing plastic from a part. That is a limit on the work it can automate.[2]
A second approach trains a model to turn sensor inputs directly into robot actions. Li says collecting suitable training data remains difficult because experience from one robot and gripper does not transfer directly to another. He hopes to combine the modular approach with learning from human demonstrations by the end of 2026, then rely more on reinforcement learning in 2027. These are research intentions, not reported deployment results.[2]
In a separate interview with The Robot Report, founder and chief executive Etienne Lacroix identified kitting as an applied research priority: gathering components into the set needed at an assembly station. His examples involve unpacking items from different suppliers and preparing them for production, rather than simply repeating the same pick from a fixed position.[3]
For manufacturers, the useful distinction is between a robot that can repeat a demonstrated skill and a system that makes another part or packaging format economical to introduce. The lab's next meaningful evidence would be a named task showing training effort, human interventions and sustained throughput against the existing pipeline. Until then, the new facility is a research commitment, not proof that those deployment costs have fallen.[1,2,3]