---
title: Hedra Founder Michael Lingelbach Takes His Stanford Robotics Research Into Physical AI With a Model Factory
description: a16z-backed Hedra, whose revenue grew 15x in a year, launches Model Factory to post-train and deploy custom physical AI models for robotics and industry.
author: Darie Nani (Editor-in-Chief)
date: 2026-11-10T16:18:54.797Z
updated: 2026-11-10T16:27:54.152Z
canonical: https://www.sovereignmagazine.com/article/physical-ai-hedra
image: https://cdn.nanimediahouse.com/hedra-michael-lingelbach.webp
categories: Artificial Intelligence, Startups
content_type: Profile
region: San Francisco
publication: Sovereign Magazine
schema_type: Article
about:
  - type: Person
    name: Michael Lingelbach
    description: Michael Lingelbach is the founder and chief executive of Hedra, the a16z-backed AI company behind the Character-3 video model and Hedra Model Factory. At Stanford University he worked on the iGibson and BEHAVIOR robot simulation platforms.
    jobTitle: Founder and CEO
    worksFor: Hedra
  - type: Organization
    name: Hedra
    description: Hedra is a San Francisco AI research company backed by Andreessen Horowitz that builds models and infrastructure for multimodal and physical AI, including the Character-3 video model and Hedra Model Factory.
    url: https://www.hedra.com
    industry: Artificial intelligence
---

As a PhD student at Stanford University, Michael Lingelbach built simulations in which robots learned to carry out household tasks. The company he founded, Hedra, is best known for AI video: its Character-3 model turns a picture, text and audio into video of a talking character, and has drawn more than 3 million users. On Tuesday the San Francisco company launched Hedra Model Factory, a service aimed at the same field as his Stanford research. It builds, runs and keeps improving custom AI models for teams in robotics, manufacturing, biology and defense.

Hedra says its revenue grew 15 times over the past 12 months. It raised a $32 million Series A in May 2025, led by Andreessen Horowitz's infrastructure fund with Index Ventures and Abstract Ventures taking part, in a round reported to value the company at about $200 million.

## Michael Lingelbach's Robotics Research at Stanford and NVIDIA

At Stanford, Lingelbach was a developer of iGibson and BEHAVIOR, simulation platforms from the Stanford Vision and Learning Lab that researchers use to train and test robots on everyday household tasks. He has also worked at NVIDIA. His work has centered on reinforcement learning, in which a model improves by trial and error against a goal, applied to visual and physical systems.

Hedra's research team includes people from Stanford, Berkeley and the University of Illinois Urbana-Champaign, with experience at NVIDIA, Meta Superintelligence Labs, Fireworks AI and Salesforce. The team has published work on generative video and embodied AI, the field of AI that controls a body or machine, at the NeurIPS, CVPR and ICLR conferences.

## Physical AI and Custom Models for Robots and Factories

[Physical AI](https://www.sovereignmagazine.com/article/what-is-physical-ai) refers to AI systems that see, plan and act in the real world, such as robots that pick and pack goods, [machines on a production line](https://www.sovereignmagazine.com/article/xpeng-iron-humanoid-robot-mass-production) and cameras that check parts for defects. These systems learn from video, sensor readings and simulations rather than mostly from text, and that data is many times larger. The simulators and training environments often need GPUs just to run at scale.

A robot that packs boxes and a camera that spots factory defects need different models, each fitted to its own task and hardware. Building and running those models can cost a company a great deal before it earns anything back. Hedra is betting that companies will pay for that work as a service rather than set up an AI research lab of their own.

## Hedra Model Factory for Post-Training and Deployment

> "The next step is the real world, where the data is video, sensors and simulation."
> — Michael Lingelbach, founder and CEO of Hedra

Customers of Model Factory, which is available from Tuesday, rent GPU capacity in Hedra sandboxes running on its inference engine, so they can scale up training without signing long-term GPU contracts. They bring their own simulation engines and data, and Hedra researchers help build the training environments and datasets, using tools such as vision-language models that caption and score data or give feedback while a model trains. For companies with no research team of their own, Hedra's researchers also set up data capture, testing and the link between training a model and running it.

Finished models run on low-latency computers, either on the customer's premises or at Hedra's nearest partner data center. That can mean distilling a world model so it runs faster than real time without losing physical accuracy, or fitting a vision-language-action model, which turns camera input and instructions into robot movements, onto a device at the edge. Once a model is in use, Hedra researchers study how it performs, retrain it and release only versions that pass the customer's own tests. Teams can also rent GPU capacity on its own or run an existing model on Hedra's inference engine.

Lingelbach describes the approach as recursive self-improvement, the idea of AI systems that get better at improving themselves. At Hedra, researchers work with autonomous AI agents that run experiments and adjust training runs, and each deployed model sends data back into the next version. "Recursive self-improvement is already changing how frontier models get built," he said, adding that Model Factory "gives robotics, manufacturing, biology and defense teams the infrastructure to post-train, deploy and keep improving their own models."

## Hedra's Results on Physics-IQ, Krea 3 and Robot Grippers

Hedra retrained its own H3 video model to produce physically consistent motion. The new version scored 34.9 on Physics-IQ, a benchmark created by researchers at Google DeepMind and INSAIT that tests whether generated video follows the laws of physics, against 30.3 for H3 Max, with similar visual quality, the company says. A distilled version rendered a five-second clip in under three seconds on AMD MI355X chips.

For Krea 3, an image model from the AI company Krea, Hedra cut the time for a 2K image edit from about 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X chips, roughly six times faster, in under a week. The original model ran at 30 sampling steps, and the faster version is a distilled model that uses 14. Krea 3 now runs on Hedra.

In a simulation study, a robot running a trained policy was given a new gripper, and its task success fell to near zero. By matching the new gripper's reach and grasp point in the view from the robot's wrist camera, Hedra brought task success back to near 100% with no retraining.

**About Michael Lingelbach**
Founder and CEO, Hedra

Michael Lingelbach is the founder and chief executive of Hedra, the a16z-backed AI company behind the Character-3 video model and Hedra Model Factory. At Stanford University he worked on the iGibson and BEHAVIOR robot simulation platforms.

## FAQ

**Q: What is physical AI?**
Physical AI is AI that perceives and acts in the real world, in robots, factory machines, inspection cameras and other hardware. It is trained on video, sensor data and simulations, and each system usually needs a model fitted to its own task and hardware, which is the work Hedra's Model Factory does for customers.

**Q: What is a physical AI example?**
Examples include a warehouse robot that picks and packs boxes, a camera on a production line that spots defects, and a robot arm that has to keep working after its gripper is changed. In Hedra's simulation study, a robot with a new gripper went from near-zero task success back to near 100% without retraining.

**Q: What is the difference between generative AI and physical AI?**
Generative AI produces content such as text, images or video. Physical AI controls or interprets machines in the real world. Hedra works in both, starting with generative video through Character-3 and now building custom models for robotics and industry.

**Q: What is Hedra Model Factory?**
Hedra Model Factory is a service that rents out GPU capacity, post-trains custom models on a customer's own data with Hedra researchers and AI agents, runs those models on low-latency hardware and keeps retraining them once they are in use.