At Rise of AI 2019 in Berlin, Ronnie Vuine, CEO of Micropsi Industries, argues that giving AI a body helps ground symbols in real experience, which is why his company connects AI to industrial robots. He contrasts traditional industrial robotics, where movements are fully predetermined and synchronized geometrically, for example welding on an Audi production line, with tasks that require perception, such as plugging in a cable, which conventional robots can only do with expensive 3D cameras and precise point-cloud model fitting. Micropsi's product, Mirai, instead lets a person guide a collaborative robot by hand through a movement while a camera records it, and sends that data to the cloud, which returns a policy for real-time, vision-guided corrective movement. He demonstrates this with a Liebherr fridge factory case, where a robot learns to find soldering points to sniff for coolant leaks, and explains that the underlying product is a general hand-eye coordination tool, not specific to that niche. Vuine explains why the company avoids reinforcement learning, relying instead on recorded real-world demonstrations, and answers audience questions on transferring the approach to other machines and on sensor choice.
Key takeaways
- Vuine says AI-driven industrial robotics has barely begun despite political rhetoric, naming KUKA and Micropsi among the few relevant players.
- Micropsi trains robots by physically guiding them through a movement while a camera records it, then sends the data to the cloud to produce a control policy.
- New collaborative robots such as Universal Robots stop on collision, making it safe for people to guide and work close to them, unlike traditional fast industrial robots.
- Vuine says reinforcement learning is impractical for factories because it needs too much real-world trial time, citing a Google X robotic grasping project he recalls as around 2015.
- Micropsi's product generalizes beyond its demo case, checking fridges for coolant leaks, to any hand-eye coordination task, such as plugging cables into a patch panel.
Detailed notes
Symbol grounding and giving AI a body
Vuine opens by connecting his talk to the symbol-grounding problem in AI: understanding language requires understanding what symbols refer to, and some researchers argue this requires bodies, since experiences like climbing a tree cannot be learned by connecting words alone. He says this is part of why he likes working at the intersection of AI and robotics, giving AIs bodies and putting them into the real world rather than having them manipulate symbols on the internet.
Why industrial robots today have almost no perception
Vuine says AI for industrial robotics has barely happened despite political rhetoric about German industrial robotics and AI competitiveness, naming KUKA and his own company Micropsi as among the few relevant players. He describes how robots in car factories, such as Audi's plant in Ingolstadt, move through fully predetermined, geometrically synchronized paths with no sensing of where a car actually is; further down the supply chain, where parts are less precisely placed, this approach breaks down and companies typically rely on humans instead. Traditional robotic solutions for tasks like plugging in a connector require an expensive 3D camera to build a point cloud and fit a precise model, unlike a human, who simply approaches and makes small corrective movements.
How Mirai trains robots by demonstration
Micropsi's product, Mirai, is a deep-learning system built on new collaborative robots, such as those from Universal Robots, that stop when they detect a collision and are therefore safe to work close to. A trainer physically guides the robot's wrist through the desired movement while a camera records it; after a couple of hours of demonstrations, the recorded data is sent to the cloud, which returns a policy enabling real-time, camera-guided corrective movement. The full setup, a robot, a force-torque sensor, a Micropsi camera, an end effector and a tablet app, is simple enough that anyone can train a complex vision-driven movement within about two hours, typically refined over two or three iterations.
Demo: checking fridges for coolant leaks at Liebherr
Vuine shows a video from a Liebherr fridge factory, where twelve soldering points on the back of each fridge must be checked with a sniffing probe for coolant leaks, a job previously done manually and prone to error as workers get tired. Automating it had reportedly been attempted for about 20 years without success; Micropsi solved it within a week of training. The robot knows roughly where each joint is from a rough model, but the last two centimeters of movement are guided live by a single RGB camera, and the system keeps tracking its target even while Vuine physically nudges the fridge or the robot during the demonstration. He notes the market for this specific application is small, around 15 installations across companies like Liebherr, LG and Whirlpool, and stresses the underlying product is a horizontal tool for any hand-eye coordination task, illustrated with further examples of plugging Ethernet cables into a patch panel.
Business model and company background
Micropsi records training data on-site at each factory to handle its specific variance, such as lighting, and returns a trained policy for that workplace rather than deploying one fixed model everywhere. Vuine says the company sells by identifying which workers remain in a factory specifically for hand-eye coordination and offering to automate that. Founded in 2014, the company spent its first two years deciding where AI work made sense to do from Germany, landed on industrial robotics and production, and now has about 20 people in Berlin plus a sales office in New York, backed by Project A.
Q&A highlights
Asked about transferring the approach to autonomous driving or harbor cranes, Vuine says the underlying hand-eye coordination skill is generic and the company plans a generic Cartesian API by year end, but that he currently declines projects like cranes to stay focused, and that autonomous driving is different because it combines many layered skills rather than one monolithic one. He says the company relies on RGB cameras rather than other sensors because the world is built for human, visual perception. Asked whether the system uses reinforcement learning, Vuine says no: reinforcement learning needs far more real-world trial time than a factory can offer, citing a Google X project he recalls as around 2015 that trained about 20 robots for four months to get a modestly impressive grasping result. Instead, Micropsi treats recorded real-world demonstrations as the training data, and transferring to new workstations requires either capturing enough variance during training or keeping new workstations physically consistent with the ones used for training.
Quotes
AI for industrial robotics really hasn't happened today.
Ronnie Vuine
The only way to make reinforcement learning work means simulation.
Ronnie Vuine
We have a strong belief that using reality as its own model is the best thing to do, and that's what we're doing here.
Ronnie Vuine
Questions answered in this talk
Why doesn't Micropsi Industries use reinforcement learning to train its robots?
According to Ronnie Vuine at Rise of AI 2019, reinforcement learning needs far more real-world trial time than a factory can offer; he cites a Google X project he recalls as around 2015, which spent four months with about 20 robots to get a modestly impressive grasping result, and says Micropsi instead trains on recorded real-world human demonstrations.
How does Micropsi Industries teach an industrial robot to plug in a cable or check a solder joint?
According to Ronnie Vuine at Rise of AI 2019, a trainer physically guides the robot's wrist through the movement while a camera records it; after a couple of hours of demonstrations, the data is sent to the cloud, which returns a policy for real-time, camera-guided corrective movement.
What sensor does Micropsi Industries rely on for its robots, and why?
According to Ronnie Vuine at Rise of AI 2019, the company relies on standard RGB cameras rather than other sensor types, because the world is designed around human, visual perception, which makes RGB the richest information source for a robot interacting with it.