mbodi to Showcase AI Robot Training at TechCrunch Disrupt 2025

Robotics Training Enhanced by Artificial Intelligence

Robots are increasingly capable of performing diverse tasks, ranging from simple box packing to complex surgical procedures. However, each specific action or process necessitates dedicated training, hindering their adaptability in dynamic, real-world settings.

Mbodi's Innovative Approach

Mbodi is developing a solution to streamline and accelerate robot training through the implementation of AI agents. The company has been selected as one of the Top 20 Startup Battlefield finalists at TechCrunch Disrupt 2025, where they will demonstrate this groundbreaking technology.

Based in New York, Mbodi has created a cloud-to-edge system – a hybrid computing architecture leveraging both cloud and local processing power – designed for seamless integration with existing robotic systems. This software utilizes a network of interconnected AI agents to efficiently gather the necessary information for rapid robot learning.

Following deployment, Mbodi’s system continuously collects data and refines its performance based on real-world applications.

How Mbodi's System Functions

According to Xavier Chi, Mbodi’s co-founder, users interact with the software using natural language prompts. Mbodi then decomposes these requests into manageable subtasks.

The company’s agent cluster effectively employs a “divide and conquer” strategy to acquire the data required for swift robot training based on the given prompt.

“The inherent complexity of the physical world lies in its infinite possibilities,” Chi explained. “The challenge arises when encountering entirely novel situations for which no prior data exists. A robust system is needed to orchestrate various models or allow for human correction and guidance in robot behavior.”

Origins and Competitive Landscape

Chi and co-founder Sebastian Peralta conceived the idea for Mbodi while working as engineers at Google. They recognized the potential for AI advancements to extend into the physical realm and identified a gap in the market for rapid robot training solutions.

Several companies, including Skild AI and FieldAI, are also focused on accelerating robot training by constructing extensive world AI models. These models rely on large datasets of real-world information to facilitate adaptation to new environments. However, Chi believes this approach is limited by the constantly evolving nature of the world.

Recent Achievements and Future Plans

Mbodi was founded in 2024, initially concentrating on picking and packaging applications. Last year, the company secured victory in an ABB Robotics AI startup competition, resulting in a partnership with the Swiss robotics organization, which was acquired by SoftBank for $5.4 billion in October.

Currently, Mbodi is collaborating with a Fortune 100 company in the consumer and product goods sector on a proof-of-concept project.

“This CPG client employs numerous individuals to pack diverse products into trays or onto shelves. The challenge is that the product mix changes daily,” Chi stated. “This constant variation makes robotic automation impractical, as reprogramming is infeasible, leaving much of the work to human operators.”

Mbodi anticipates broader deployment of its software beginning in 2026.

“Our goal is to develop a practical, deployable solution,” Chi emphasized. “We are not a research laboratory and prioritize creating a reliable, production-ready system.”

To learn more about Mbodi’s technology and explore valuable workshops, attend TechCrunch Disrupt, taking place October 27 to 29 in San Francisco.

This article has been updated to accurately reflect Xavier Chi’s professional title.