FAIR Embodied AI

Embodied AI is the discipline of building agents that don’t just talk — they move, look, reach, navigate, and act. At FAIR (Meta’s research org), I led design across multiple teams working on the layer that turns research models into systems people can actually use, across glasses, headsets, and robotics.

The two headline projects are Motivo and SIRo. Motivo is the model that gives a humanoid character a usable repertoire of motion. SIRo is the project that gives a robot enough social intelligence to share a room with a person.

I love robots. The whole point of this stretch of work was to make them a little more useful and a lot less weird.

Most large models can describe the world. Few can act in it. Motivo is a behavioral foundation model for humanoid control — trained on a wide library of motion data so that a digital character (and eventually a physical one) can be prompted to walk, sit, reach, dodge, balance, recover from a stumble. Without per-task training. Without scripted animation.

The design work sat at the seam between the research and the people who would eventually use it. Multimodal prompts (gesture, motion, reward inputs). Interaction patterns that don’t exist yet. Visualizers and review tools so a non-PhD could sanity-check what the model was doing. The team needed product-grade UX before there was a product.

What it’s for: avatars in social VR that move like people, not mannequins; humanoid simulation for training data; eventually, the motion layer that lets a real robot do real chores. The interesting part isn’t the demo — it’s the substrate.

Most robots fail not because they can’t do a task, but because they can’t share a room. They walk through a doorway you’re standing in. They reach for an object you’re holding. They don’t notice a child or a pet or a person carrying something heavy. SIRo (Socially Intelligent Robots) is the long arc of getting machines past that ceiling.

The design work covered three loops: perception (the robot’s sense of where humans are and what they’re doing), intent (what it’s about to do, communicated before it does it), and handoff (when the human takes over, when the robot offers help, when neither). Designing those loops is mostly designing for moments of mutual surprise — what does it look like when the robot and the person disagree, and how does the robot back off gracefully?

Public-facing work covered home assistance scenarios: setting a table, fetching a tool, holding a door, navigating a hallway in a house where toys, children, and laundry baskets are in motion. The point isn’t a demo of the perfect day. It’s a robot that doesn’t make the messy day worse.

Built and led four teams across the embodied-AI surface area, each pairing design with a partner research org. The work moved between research lab and product surface, sometimes weekly.

Today’s AI is in a cave. It’s trained on images, text, audio, video — flat data, with no body, no consequences, no reward function tied to physical reality. An agent that drops a glass and watches it shatter learns something a chatbot never can. Embodiment isn’t a moral upgrade to current AI. It’s an architectural requirement.