By working with embodied AI, I keep getting pulled deeper into how human interaction really works — with objects, with space, with each other. Even a simple task carries an enormous amount of backend processing more of us never notice. A few takeaways from watching Figure’s latest Helix 02 demo:

1. The invisible decisions behind “just making a bed”
   Watching Helix 02 make a bed, I found myself asking questions I’ve never consciously asked my own mornings: What made it place the pillow first, before unfurling the duvet? How does it decide where to hold the fabric when stretching it out? How does it even measure “done”? I make my bed the same way every day without thinking about any of this — or more accurately, I probably am thinking about it, just without being aware of the thinking. If a robot can eventually absorb tasks like this, what would we actually do with the time and attention we get back?

2. Is efficiency even the right goal?
   A friend once asked me, “Have you ever tracked whether every decision you make is for efficiency?” We both work in fields built around making life more efficient, so the question stuck with me. My own life is full of counterexamples — I know I should clear surfaces before vacuuming, and I usually don’t, sometimes out of laziness, sometimes because there’s nowhere else to put things. Which raises a real design question: should a robotic system always optimize for the most efficient way to complete a task? Efficiency isn’t a single, universal value for humans, and it isn’t always achievable — so why would we assume it’s always the right target for a robot either?

3. Choosing how to complete a task
   In the clip, Helix 02 closes a book resting on a pillow and sets it on the nightstand — one hand closing the book while the other simultaneously clears space to place it. Two independent actions, running in parallel. There’s more than one way to close a book, more than one way to make room for it, and the system had to choose one. The choice cannot be neutral — it’s shaped by the robot’s own physical constraints, the object’s constraints, and the environment around it.

4. Defining “complete” in the physical world
   Even something as small as “close the book” hides a cascade of possible definition of done:
   a. Close it and leave it exactly where it was
   b. Close it and put it where the user alwasy keeps it
   c. Close it and return it to its place on the shelf
   d. Actually — what if the user wanted it left open, to pick up where they left off?
   None of these is objectively correct. Each is a different bet about what the human actually wants, and getting it wrong doesn’t look like a bug — it looks like the robot didn’t understand you.

5. Two robots — and what happens at three?
   It seems like making a bed isn’t a one-robot job yet, so Helix 02 does it as a team of two. What’s striking isn’t just the coordination — it’s that there’s no shared planner and no communication channel between them. Each robot reads the other’s intent purely maybe via the software system or from the motion, the way two people folding a sheet don’t need to talk about it. At three or more, then, does it start requiring something like assigned roles — a persona or set of attributes per robot — just to avoid two robots claiming ownership or the same task?

Reflecting on the backend behind such ordinary human behavior makes clear how hard these “simple” tasks really are for a mechanical system. Should a system eventually try to mimic the entire human thought process to get the job done? Or is the real work upstream of that — figuring out which of these invisible decisions the system should make silently, and which ones it should check with the human first? That’s the same question I ask when I’m building an interaction design framework: not just what can system do, but what does the user actually want it to achieve?