Worked with · 11 of 12Humanoid robotsapptronik.com

Apptronik

I collaborated with the Apptronik team on work around embodied AI, humanoid deployment and the gap between a robot that performs well in a controlled demonstration and one that can work reliably in a real environment.

Relationship: Worked with
Company: Builds Apollo, a humanoid robot for industrial work
Format: Collaboration
Focus: Embodied AI and humanoid deployment
Covered: Learning from variation, deployment data, reliability, working around people, hardware and models

Never quite static

A large part of my work focused on that transition. Industrial environments are repetitive in some ways, but they are never completely static. Objects move, people enter the workspace, layouts change and tasks that look identical at a high level can require slightly different manipulation every time.

Fig. 1Repetitive, never static

Objects move
People enter the workspace
Layouts change
The same task, slightly different manipulation

Learning from variation

I spent time looking at how Apollo could learn from those environments without treating every variation as a new task. That meant thinking about the relationship between perception, task planning, manipulation and the physical limits of the robot itself.

Fig. 2Four parts of one behavior

PerceptionWhat the robot sees
Task planningWhat it decides to do
ManipulationHow it handles the object
Physical limitsWhat the robot itself can do

Data

Data collection was an important part of the work. Humanoid systems need experience from real tasks to improve, so deployment also becomes a source of training data. I looked at what information was useful to collect, how teleoperated and autonomous behavior could be compared, and how failures could be turned into examples for the next training cycle.

Fig. 3Deployment as a source of data

01Real tasksThe robot at work
02CollectionTeleoperated and autonomous
03FailuresCould become examples
04Next cycleTraining, then back to work

Reliability

Reliability mattered more to me than isolated success. A robot completing a task once tells you much less than the same system completing it hundreds of times while handling small changes in object placement, timing and surrounding activity. I was interested in where performance started to degrade and whether the cause came from perception, planning, control or the physical interaction itself.

Fig. 4Once, and then hundreds of times

Once, in a demoHundreds of runs, small changes each timeWhere it starts to slipFirst runRepeated use

Schematic, not data

The interesting point is where performance starts to slip, and whether the cause is perception, planning, control or the physical interaction.

Around people

Human interaction was another part of the problem. Apollo is designed for workplaces that were built for people, which means the robot has to share space with existing workers and equipment. That makes predictability, safe motion and recovery behavior part of the system design rather than something that can be added after the core task works.

Fig. 5Designed in, not added later

Predictability
Safe motion
Recovery behavior

Hardware and models

I also looked at the feedback loop between hardware and embodied AI. Better models can improve what the robot understands and how it chooses an action, but the training data still comes through a physical platform with its own sensing, dexterity and mobility constraints. Changes to one side affect what is possible on the other.

Fig. 6Two sides of one loop

Models

What the robot understands
How it chooses an action

The platform

Sensing
Dexterity
Mobility

The training data comes through the platform, so a change on one side changes what is possible on the other.

What I took from it

The work gave me a better way to think about humanoid robotics. The difficult part starts after a capability works in a demo. The real question is whether the same behavior survives repeated use, unfamiliar situations and the ordinary messiness of a workplace.

Materials

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

Screenshot of Welcome to Robot Park: where Apptronik’s Apollo goes to work, Apptronik · June 2026

Screenshot, Oct 5, 2026 Apptronik · June 2026 Welcome to Robot Park: where Apptronik’s Apollo goes to work Fleets of Apollo 2 robots collecting real-world data across Robot Park sites, in a research partnership with Google DeepMind.