Figure AI just released Helix 02, trained on over 1,000 hours of human motion data. Their robot can unload a dishwasher, walk across a kitchen, stack dishes in cabinets, and start the dishwasher again. Four minutes of continuous autonomous behavior. No human intervention. They replaced 109,504 lines of hand-engineered C++ with a single neural network trained on human movement.

It is impressive work.

And it has almost nothing to do with whether you'd want that robot in your house.

When someone interacts with a robot for the first time, they read its body language before evaluating its grip strength or path planning. They do this automatically, the same way they'd read a person who just walked into the room. Within milliseconds, before any conscious thought, they've already decided: do I trust this thing?

While utility is becoming table stakes, trust is built through posture and timing.

The Disney Insight

Disney Research recently published a paper on building a physical Olaf robot. The entire project is organized around a single word: believability. They spent their engineering budget on reducing the sound of footsteps, because harsh robotic impact noise breaks the illusion of life. They built a thermal-aware control policy so the robot could operate for long stretches without overheating, because a character that needs to cool down every five minutes stops being a character. They optimized for the way the robot feels to be around, and every technical decision flows from that priority.

Disney understands something the robotics industry mostly doesn't. People evaluate robots the way they evaluate people. The evaluation is subconscious and nearly instantaneous. You can build the most capable manipulator on the planet, and if its idle posture reads as "dead" or its head movements feel twitchy, people will not want it near them.

Where the Data Comes From (And Doesn't)

The current playbook for training robot behavior is teleoperation. A human operator controls the robot remotely, demonstrating tasks, and the system learns from those demonstrations. This works well for manipulation: pick up the cup and move it. The operator's intention maps cleanly to the robot's action.

But teleoperation captures task intent, not social intent. When an operator is focused on guiding a gripper toward a dish rack, they are not simultaneously performing the subtle weight shifts and head tilts that would communicate attentiveness to another person in the room. Teleop dictates the robot's physical actions without addressing its non-verbal communication.

The second obvious escape route is control theory. Just try to program the non-verbal behaviors by specifying the angles and the trajectories. But this runs into the same wall that Figure themselves acknowledged when they replaced over a hundred thousand lines of hand-engineered locomotion code with a learned system. While hard-coding a nod is certainly possible, people do it all the time, crafting the timing of a nod that reads as agreement versus a nod that reads as dismissal is not. The difference between those two nods is maybe forty milliseconds and a slight change in acceleration profile. The combinatorial space of contextually appropriate non-verbal responses is too large to specify with rules. It has to be learned.

Then there's simulation. You can simulate a robot unloading a dishwasher with reasonable fidelity. You can model rigid body dynamics, contact forces, friction. Sim-to-real transfer works for locomotion because the physics of walking are well understood and the success criteria are clear: did the robot fall over or not?

Non-verbal communication doesn't have clear success criteria in simulation. There is no loss function for "did that head tilt feel like curiosity or a servo glitch." The micro-hesitations and the way a human shifts weight before speaking live in a regime of subtlety that simulation cannot generate, because simulation doesn't know what to optimize for. The ground truth lives in human bodies.

Current methods of simulation, teleoperation, and control theory fail to capture or specify these subtle behaviors. The only remaining path is large-scale human motion data, extracted from real human behavior and retargeted onto robot embodiments.

We tested this hypothesis on our robot, JoeBot. We generated or captured motion and used our universal retargeter Mimik to put the motion, subject to physical constraints, onto this wheeled robot. The result is a really compelling character, with utility, that can communicate naturally.

The Scale Problem

Figure needed 1,000 hours of motion data for a single form factor doing a single class of task. That's a serious data collection effort, and it produced remarkable results for locomotion and manipulation.

Non-verbal communication requires exponentially more data.

Active listening alone is a complex behavior. The slight forward lean, the small nods at irregular intervals, the gaze tracking, the postural mirroring. An affirmative response carries a distinct faster, more vertical motion signature with distinct hand and arm involvement. A warning changes everything about the body's spatial relationship to the listener, the forward projection of the head, the tension in the shoulders, the speed of the gesture. And personality, the ambient motion style that makes one entity feel warm and another feel cold, is the most data-hungry problem of all, because it modulates every other behavior continuously.

Each of these is a distinct motion vocabulary. Each requires its own training distribution, they compound, the robot needs to transition between them as fluidly as the human interacts. The motion data required to make a robot socially fluent will dwarf what's needed for task execution by orders of magnitude.

Every company building humanoid robots will eventually need millions of hours of this data. And almost nobody is collecting it.

Movement from Pixels

I recently argued that video is the largest unrefined resource in human history. Billions of hours of humans in motion, recorded from every angle, in every context, performing every social behavior that matters. Video is data to be mined, not just content to be consumed.

Computer vision has reached the point where high-fidelity movement data can be extracted directly from the pixels of ordinary video. High-fidelity movement data can now be extracted directly from ordinary video pixels, bypassing motion capture suits and expensive production pipelines. The physics of how humans actually move through space and interact with each other is already recorded. It just needs to be read.

A universal retargeting system completes the bridge. Motion extracted from a human in a video can be mapped equality well onto digital avatars, game characters, or humanoid robots. The same motion data, the same human intent, expressed through whatever body needs to express it.

This is what we've built. JoeBot can perform active listening, give affirmative responses, issue warnings, and express personality through movement, all driven by human motion data captured and retargeted through our system. These movements originated in human bodies, carrying the micro-variations and timing that make them legible to other humans. No hand-animation. No procedural generation.

Why This Matters Now

The robotics industry is racing to demonstrate task capabilities like sorting pills or unloading dishwashers. These are important milestones, and the companies achieving them deserve credit.

But task capability alone is insufficient for robots living alongside people. The gap between a robot that can do chores and a robot you'd actually want in your home is filled almost entirely with non-verbal communication. The posture it holds while idle. The way it turns toward you when you speak. The speed at which it responds to an unexpected sound. These details determine whether a machine feels like a tool or a presence.