Organization: Mbodi AI
Country: U.S.
Website: mbodi.ai
Year Founded: 2024
Number of Employees: 1-10
Innovation Class: Technology
For all of the attention on the latest artificial intelligence models, relatively few companies are applying them to robotics. Mbodi AI said it enables robots to learn and adapt to new tasks instantly through natural language commands.
“We’re augmenting the industrial robotics experience through cutting-edge generative AI that fulfils evolving customer needs,” said Sebastian Peralta, a former Google engineer and co-founder and CEO of Mbodi.
The New York-based company uses transformers and an agentic infrastructure to help robots learn tasks in minutes through written and spoken interaction plus demonstration.
Mbodi’s core innovation is MbodiOS, a modern, cloud-to-edge operating system for real-time agent orchestration. The system uses a hybrid architecture that combines generative AI with classical robotics techniques like path planning and object detection, enabling continuous real-time adaptation and learning.
Robots can execute actions in under 0.5 sec., allowing real-time modifications to robot behavior without requiring complex programming or expensive training, said Mbodi. Instead of relying on predetermined scripts or offline model updates, the company said it delivers continuous, “on-the-fly” adaptability.
“We believe an agentic architecture will enable one application in a domain,” said Marc Segura, president of ABB Robotics, which recognized Mbodi a winner of its 2024 ABB Robotics AI Startup Challenge. “For example, rather than train a neural network in electronics assembly, we want to enable the manufacturer to tell a robot to bend wires and put them on a specific object.”
“It’s also important to have vision plus tactile sensing for end-to-end training of robot skills with imitation,” he said. “Rather than build a robot that can do all the things to clean a hotel room, we’re looking at an atomic approach.”
Mbodi said that it expects vision language action models (VLAs) to eventually help robots generalize when used with existing technologies such as machine vision and cloud computing.