On September 9th, Vention announced the launch of its new Physical AI Lab. Spanning Physical AI and industrial robotics, the goal is to advance robotic manipulation from initial research to scalable systems for today’s factory floors.
The lab’s agenda covers industrial data collection, applied manufacturing use cases, and post-training foundation models for robotics. Vention’s Content Editor, Mariah Moore, sat with Dr. Jimmy Li, Director of Physical AI, to discuss the lab’s scope of work, predicted outcomes, and the near future of Physical AI in industrial automation.
Dr. Jimmy Li is the Director of Physical AI Research at Vention, where he develops and applies advanced AI technologies to drive innovation in industrial automation. He plays a key role in shaping intelligent automation solutions, and his work bridges theoretical research and real-world impact. Jimmy holds a Bachelor of Arts & Science and a PhD in robotics from McGill University.
Vention: What made now the right time to open a dedicated Physical AI lab, rather than continuing research inside existing teams?
Dr. Jimmy Li
“Right now, we’re gaining traction in Physical AI with more demand and more customers. We’re seeing an upswing in requests to deploy Physical AI cells across industries, so it makes sense to add more firepower to the Physical AI playbook.
I also think there’s a lot of activity in the research community. Many new technologies are coming out that could reshape manufacturing, and we want to play a part in shaping it.”
Vention: The framing of the lab is around bridging AI research and scalable automation deployment. Where does that bridge actually break down for companies trying to deploy today?
Dr. Jimmy Li
“So there are two approaches to deploying Physical AI. It’s helpful to understand these two categories first, because they each have their own challenges.
The modular approach is where perception and motion planning are split into two different components. Usually, the perception model will output some human-interpretable representation of the scene. It might output, “Okay, there’s an object at this location,” and then the motion planner grabs it. That is the approach we’re using for most of our deployments right now.
The challenge here is the skill ceiling. There’s more analytical and manual programming involved, and it’s very hard to program the robot to perform dexterous tasks like peeling plastic off of a part. On the positive side, these pipelines use a lot of vision foundation models as building blocks and are starting to become zero-shot and quick to set up.
The second approach is the robot learning approach, which we often refer to as sensor-to-action policies. That’s where a single model takes sensor data as an input and directly outputs low-level robot actions. When we talk about ‘post-training foundation models’, that also falls into this bucket.
These methods have a much higher skill ceiling. You can have a human teach it to do very dextrous things, but the challenge is collecting a lot of data to fine-tune a model on the specific hardware used during deployment. Under both approaches, there is a scalability gap.”
Vention: Thanks for breaking that down. Would the team in the new lab be prioritizing that second sensor-to-action method because of the skill ceiling of the modular approach?
Dr. Jimmy Li
“Both methods are equally important, because they’re actually intertwined. If we have a good modular pipeline, it can be used as a starting point to help the sensor-to-action policy.
My hypothesis is that we can use the modular pipeline to get a system up to 60-70% of the performance we need. Then we use the sensor-to-action policy to learn off of that to get to 100%. Essentially, the systems are two sides of the same coin.”
Vention: I hadn’t realized they were so intertwined. What do you anticipate will be some of the key outcomes or findings from the work in the lab, aside from confirming that hypothesis?
Dr. Jimmy Li
“The high-level goal is to make onboarding tasks much easier for robotics manufacturing. For now, we’re focusing a lot on pick-and-place tasks, and there are many variants, like bin picking and kitting. We also want to push the skill ceiling for robots to perform highly dexterous actions.
To facilitate part onboarding, we want to have an efficient way to bootstrap the robot system by using our modular pipeline and few-shot foundation models. I also believe reinforcement learning is an important piece. It’s where the robot tries different things on its own and potentially achieves superhuman performance with enough practice.
Today, when we deploy a cell, it is 100% done with the modular pipeline, but I hope that by the end of this year, we’ll start to be able to mix the modular pipeline with Learning From Demonstration (LFD). Later on in 2027, we’ll hopefully reduce the amount of teleoperation and have reinforcement learning do most of the heavy lifting.”
Vention: The research side, according to the release, spans applied manufacturing use cases, data collection, and post-training foundation models. Which of those three is the biggest challenge right now?
Dr. Jimmy Li
“I think post-training foundation models are definitely the hardest one. As I mentioned, it typically requires a lot of data collection, so making it scalable is still an open area of research.
There are datasets out there, but if that data is collected on a different embodiment, with a different robot and a different gripper, they wouldn’t directly transfer to my own robot and my own gripper. That’s why typically for every deployment, you need to collect data manually. We need to make that easier.”
Vention: And what would you say makes Vention well-positioned to tackle these areas of research versus other key players in the industry?
Dr. Jimmy Li
“Just the fact that we’re in factories, I think, is such a big factor. We have a very good partnership with NVIDIA, and one of the reasons that we’re a valuable partner is that we have a channel to the customer. We see the problems on real factory floors, and that’s really important for focusing our research on the right areas.
Also, if we’re thinking in general about deploying a cell, we have a hardware platform and all the levers of an end-to-end solution. That’s also very important.”
Vention: We wanted to touch on Dr. Pineau’s role as an external technical advisor. What makes her a key addition to the lab?
Dr. Jimmy Li
“I’ve known Dr. Pineau’s work for a long time. She has a Ph.D. in Robotics from Carnegie Mellon University, with experience ranging from Meta to Cohere to the board of the Journal of Artificial Intelligence Research.
She’s an expert in developing new models and algorithms for learning and planning in complex domains. Her work is very impressive, and she’s an exciting addition to the team, especially in an advisory role.”
Vention: She sounds like a great addition. Is there anything else people should know about the launch of the Physical AI Lab?
Dr. Jimmy Li
“I think there’s a lot of interest when we talk about bridging the gap between research and the real floor. As you know, we encourage clients to send us parts, and when we can work with real pieces and use cases, the team can create a feedback loop to build their solution more effectively. I think that’s a big aspect of the lab that we’re excited about.
We’re always talking to real engineers out on the floor. Balancing research, development, and deployment is important, and keeping close contact with the specialists in the field is the only way to make that happen.
There’s a lot of great Physical AI research being done in Canada. Montreal is a hub with many great universities and research labs actively working in this area. We’re always looking for ways to connect and collaborate with the academic community.”
To learn more about how Vention Physical AI could be applied to your factory floor, send us your parts or talk to an expert.
Click to read more about Vention’s new Physical AI Lab.
