Alina Schanz

Worked with · 10 of 12Humanoid robotsfigure.ai

Figure

I collaborated with the Figure team on work around physical AI, real-world humanoid deployment and the systems needed to turn learned behavior into reliable work.

Relationship
Worked with
Company
Builds general-purpose humanoid robots and the Helix model that runs them
Format
Collaboration
Focus
Physical AI and real-world humanoid deployment
Covered
Long tasks, deployment data, human data and robot learning, reliability at scale

A shift, not a demo

A large part of my focus was the gap between a capability that works in a controlled setting and one that can survive repeated use in a factory or logistics environment. Once a humanoid becomes part of an operating workflow, success is measured across entire shifts rather than individual demonstrations.

Fig. 1Two ways to measure success

A

A demonstration

  • One run
  • A controlled setting
B

An operating workflow

  • Entire shifts
  • A factory or logistics floor

Longer tasks

I spent time looking at how perception, manipulation and movement interact during longer tasks. With Figure’s Helix architecture, the robot has to interpret what it sees, decide how to act and coordinate its body continuously as objects and its own position change. Errors accumulate differently over a long sequence than they do during a single pick-and-place action.

Fig. 2What runs continuously

  1. 01SeeInterpret what it sees
  2. 02DecideHow to act
  3. 03MoveCoordinate the body
  4. 04ChangeObjects and its own position move
Over a long sequence the loop repeats, and small errors carry into the next turn of it.

Deployment data

Deployment data was especially important. Repeated work exposes cases that are difficult to reproduce in a lab: slightly different object placement, changing lighting, worn containers, people entering the workspace and small variations in the way a task is presented. I looked at how those failures could be classified and fed back into training instead of being treated as isolated incidents.

Fig. 3What a lab rarely shows

  1. Slightly different object placement
  2. Changing lighting
  3. Worn containers
  4. People entering the workspace
  5. Small variations in how a task is presented
How such failures could be classified and fed back into training instead of being treated as isolated incidents.

Human data

I was also interested in the relationship between human data and robot learning. A general-purpose humanoid cannot rely on collecting a new set of robot demonstrations for every environment it encounters. Figure has been moving toward learning from large amounts of human behavior, which creates a different research problem around what transfers from human motion and decision-making into a robot with different physical constraints.

Fig. 4Where the experience comes from

A

Robot demonstrations

  • A new set for every environment
  • Hard to scale for a general-purpose humanoid
B

Human behavior

  • Available in large amounts
  • Has to transfer to a body with different physical constraints

BMW as a reference point

The BMW deployment was a useful reference point for this kind of work. Figure 02 moved from initial testing into an active assembly line and eventually accumulated more than 1,250 hours of runtime while handling over 90,000 parts. That scale makes it possible to study reliability, recovery and performance drift in a way that short demonstrations cannot.

Fig. 5Reported by Figure: Figure 02 at BMW

  1. 1,250+hours of runtime
  2. 90,000+parts handled

Figures as reported by Figure

Enough repetition to study reliability, recovery and performance drift.

What I took from it

The question I kept coming back to was how much of a robot’s capability survives contact with the real world. For me, that meant looking beyond whether a task could be completed once and paying more attention to repeatability, failure recovery, useful training data and whether the same learned behavior could transfer to the next environment.

Fig. 6Beyond completing a task once

  1. Repeatability
  2. Failure recovery
  3. Useful training data
  4. Transfer to the next environment

Materials

Public pages about the company and the programs around this work.

Screenshot of Figure × Brookfield strategic partnership, Figure · September 2025Screenshot, Oct 5, 2026 Figure · September 2025 Figure × Brookfield strategic partnership On building a large real-world dataset for humanoid pretraining with Brookfield.