You don't rewrite a new hire's code. You show them. Mbodi wants robots taught the same way.
Here is a small, unglamorous truth about factory automation that Mbodi AI has organized its entire company around: the robot arm is the affordable part. What costs money - real money, the kind that keeps mid-sized manufacturers out of automation entirely - is telling the arm what to do. Reprogramming an industrial robot for a new task has traditionally meant weeks of work by a specialized engineer, per task, per line. Robots are wonderful at repetition and hopeless at change, and most actual manufacturing lives in the space between.
Mbodi (say it "embody," which is rather the point) is a New York company founded in 2024 that proposes to close that gap with language. Its pitch, delivered by co-founder Sebastian Peralta, is unusually blunt for a robotics startup: "We're building an embodied AI platform that lets anyone teach robots new skills - just by talking to them. No code. No engineers." You describe the task. You show the robot a quick demonstration. It gets to work, in production, in minutes, and adapts as often as you need it to.
Underneath that friendly promise sits a genuinely hard engineering problem, which is the interesting bit. Large language models are fluent, confident, and occasionally wrong - a combination that is charming in a chatbot and expensive on an assembly line, where "occasionally wrong" means a dented part or a stopped line. So Mbodi doesn't just bolt a language model onto a robot and hope. Its platform fuses generative AI with agent orchestration and symbolic reasoning, layering structure on top of the model's fluency so that natural-language intent turns into precise, reliable, repeatable robot actions.
The architecture runs cloud-to-edge and, importantly, hardware-agnostic. A universal agent framework coordinates perception, reasoning, planning, and control across whatever robot you happen to own. And there is a quietly powerful feature buried in that design called fleet learning: teach one robot a skill, and every other robot on the network inherits it. Show it once; deploy it everywhere. That is the sort of leverage that turns a clever demo into a business.
The company is starting where the leverage is highest and the risk is lowest - pick-and-place, the humble task of moving an object from here to there. It is boring, which is exactly why it is the right beachhead. High-mix, low-volume production is full of pick-and-place variations that never justify weeks of engineering, and that dead zone is precisely what a language-first system is built to serve.
The workflow is deliberately shaped like teaching a person, not programming a machine. That framing is the product.
Describe the task in plain natural language. No scripting, no robot-specific syntax.
Give a short demonstration. The system watches, perceives, and infers the intent.
Generative AI and symbolic reasoning turn intent into precise actions - in production within minutes.
Fleet learning transfers the new skill across every robot on the network. Adapt on the fly.
“A world where robots are as easy to teach and adapt as software.”
Both co-founders spent time on Google Public DNS - the 8.8.8.8 service a large slice of the internet quietly depends on. That background matters more than it sounds: it is a job about reliability at enormous scale, and reliability is the whole game when you point AI at the physical world.
A physicist, roboticist, and deep-learning researcher. Triple major in Electrical Engineering, Computer Science, and Physics from UPenn, with graduate robotics research at the university's GRASP Lab. Formerly a core engineer on Google Public DNS.
Former tech lead of Google Public DNS (8.8.8.8), one of the world's most critical internet services, with work covered in The Register. Holds bachelor's and master's degrees in Electrical & Computer Engineering from UIUC.
Cloud-to-edge system that turns natural-language instructions and short demos into precise, reliable robot actions - running in production in minutes and adapting on demand.
Hardware-agnostic coordination of perception, reasoning, planning, and control across any robot - with fleet learning so a skill taught once transfers everywhere.
Open-source Python SDK (on PyPI as mbodied) to drop transformer models into a robot stack in a few lines - with HuggingFace, Gymnasium, Ollama & OpenAI-compatible support.
Who is it for? Manufacturers and robot operators - especially the high-mix, low-volume shops where reprogramming has always been the bottleneck. And developers, who can start today by reading the open-source toolkit before they ever talk to sales.
Mbodi AI is founded in New York by Sebastian Peralta and Xavier (Tianhao) Chi.
Wins ABB Robotics' global AI Startup Challenge, opening a joint commercialization partnership with the world's leading robot maker.
Unveils an updated company logo - "new look, same mission" - built around real-time embodied skill learning.
Launches publicly on Y Combinator as an embodied AI company for industrial robotics. Named a 2025 RBR50 Robotics Innovation Award recipient.
Selected to demonstrate training a robot live with AI agents at TechCrunch Disrupt 2025.
Mbodi is early-stage and Y Combinator-backed, with reported seed funding of roughly $130K (approx.). The more telling asset is the ABB relationship - a commercialization partner that already sells robots to the world.
Mbodi shows its work in public - launch talks, live robot-training demos, and an open codebase. Start here: