Mbodi will show how it can train a robot using AI agents at TechCrunch Disrupt 2025

Robots can be programmed to perform a variety of tasks, such as packing boxes and even performing surgeries. But each individual movement or task requires its own specific training process, making it difficult for robots to adapt to real-world scenarios.

Mbodi wants to make training robots easier and faster using AI agents. The company will showcase this technology as one of the Top 20 Startup Battlefield Finalists at TechCrunch Disrupt 2025.

New York-based Mbodi built a cloud-to-edge system, a hybrid computing system that uses both cloud and local computing power, that is designed to integrate with existing robotic technology. The software relies on a large number of AI agents communicating with each other to gather the necessary information so a robot can learn a task faster.

Once implemented, Mbodi will collect data and learn from his real-world examples.

Xavier Chi, co-founder and CEO of Mbodi, told TechCrunch that users control the software in natural language, and Mbodi breaks the request down into smaller subtasks. Mbodi’s cluster of agents essentially divides and conquers the task of gathering the necessary information to quickly train the robot on the prompt.

“The tricky thing about the physical world is that there are infinite possibilities,” Chi said. “Any time you can invent something completely new, you don’t have data, that’s a problem in the physical world. We always need a system where you can orchestrate different models or have someone correct a robot and tell it to do certain things in a certain way.”

Chi said he and co-founder Sebastian Peralta got the idea for the company while he was working as an engineer at Google. Although they weren’t working on robotics, they both came to realize that advances in AI were moving into the physical world and despite an increase in physical AI, there still wasn’t a great way to quickly train robots.

Many companies, like Skild AI and FieldAI, want to help make training robots faster by building AI models for the entire world with enough real-world data so they can more easily adapt to new environments. Chi said philosophy just doesn’t work with the rate at which the world is constantly changing.

Mbodi was launched in 2024 with a focus on picking and packaging. The company won an AI startup competition from ABB Robotics last year, which earned them a partnership with the Swizz robotics organization that was acquired by SoftBank for $5.4 billion in October.

Now the company is working with a Fortune 100 consumer and product company on a proof of concept.

“For the CPG customer, they have a lot of people, they package different products of their brand in a container or on a shelf, the problem is it changes every day,” Chi said. “That’s why it’s impossible to put robots there. Reprogramming these robots is just not possible, there are still a lot of people doing that work.”

Mbodi hopes to use his software more widely in 2026.

“We want to build something that works, that can actually be deployed,” Chi said. “We are not a research laboratory; in that respect we do not want to be a research laboratory. We want to put something into production that works reliably.”

To hear from Mbodi firsthand and see dozens of additional pitches, attend valuable workshops, and make the connections that drive business results, Go here for more information about this year’s Disruptheld October 27-29 in San Francisco.