Meet Mbodi, Automate’s 2026 Startup Challenge winner

Several startups are racing to build massive models to learn adapt to new tasks. Xavier Chi, co-founder at Mbodi, doesn’t think that strategy will actually work.

“The physical world [offers] unlimited possibility, it's a big challenge to really scale that into a solution that works reliably,” Chi told A3 on a recent episode of Automate Live. ”When the model is trained, it's like a black box, and when it fails in production, we have no idea what's going on in there, and hope we should collect more data and train the system so that maybe it can fix the error it just did, and that's a big challenge.”

Chi’s startup, Mbodi, takes a very different approach to robot training. Instead of a one-size-fits-all approach, they're developing small models and agentic AI.

The New York-based startup prompts software with natural language. AI agents then break down each ask into smaller tasks then work together to source the information needed to train the robot on the task at hand. The company is currently working with collaborative robots in industrial use cases first.

Mbodi recently got further validation on its idea and won Automate’s 2026 Startup Challenge for early-stage companies, beating out nine other finalists in the eyes of judges from tech heavyweights like Nvidia and Microsoft.

“Automate is a great show,” Chi said. “It's our favorite show, actually, one of the biggest in the whole robotic space, and we just thought we should go on stage and pitch our product or technology and see how people react to it. We got some really good feedback from people from all across different domains, so that's a really great experience.”

ROI Calculator

Discover the potential cost savings of robotic automation over a 20-year system life

This calculator compares your current manual labor costs against the total cost of owning and operating a robotic system over its 20-year lifespan.

EXPLORE TODAY

Chi and his co-founder Sebastian Peralta got the idea for the company when they were working as engineers at Google. While not focused on robotics specifically, noticed that, despite advancements in AI and in physical AI, there still wasn’t a great way to train robots.

Mbodi is focused on deploying its tech as much as it can over the next year, Chi said. He added that they don’t want to be known as a research company but rather as a tool that companies can start implementing today.

“To scale it up, we need to make sure about the flexibility, capability, the speed, throughput, and cost, so it makes sense for the customer, and we have to maximize all of those three aspects to make sure it really works,” Chi said. “We don't want to, for example, stay in research mode for many, many years. I don't think that's the right way to do it. The right way is to make sure we listen to our customers.”

The startup launched in 2024 with a focus on picking and packing. They’ve since been able to get that task to  99.96% accuracy, Chi said. The company has evolved its tech since it started. It originally was a cloud to edge system but after customer requests, the company moved its vision system to edge devices so the whole system can run locally.

Mbodi has seen a lot of demand especially from the pharmaceutical space, Chi said, and the company is currently working on a pilot with two of the largest companies in that industry, although Chi couldn’t share any details.