I presented at DevelopersIO 2026 Tokyo with a technical stack and real project demo of Physical AI, the AI that has come down to the field #devio2026

Hello.

My name is Takaaki Tanaka from the Smart Factory Team in the Manufacturing Business Technology Division.

This blog covers the content I presented at DevelopersIO 2026 Tokyo under the title "AI Descending to the Shop Floor: The Physical AI Technology Stack and Real Project Demo."

Presentation Slides

Key Points from the Presentation

Three Hurdles That Dropped All at Once Over the Past Year

Physical AI has become much easier to get started with over the past year or so. I believe the following three hurdles have been lowered.

• A place to experiment (simulation): NVIDIA Omniverse and Isaac Sim allow you to reproduce a real site in virtual space, physics and all. Isaac Sim has been released as open source (Apache 2.0).
• Motion data (pre-trained model of human movement): Kimodo, a model trained on approximately 700 hours of human motion collected via optical motion capture, has been released.
• Body control (foundation model for whole-body control): GEAR SONIC has been released. A single model can follow whole-body movements such as walking, running, and crouching.

The hardware is expensive? …Yes, it is.

Technology Stack: 5 Layers from Words to Robot Joint Movement

Here is an explanation of the technology stack used. To make it easier for me to understand, I organized the whole thing into five layers.

Isaac Sim (L2: Test in Advance)

What we actually interact with in the robot context is Isaac Sim, built on top of Omniverse. The features of Isaac Sim are as follows.

• Released on GitHub under Apache 2.0 and can also be installed via pip
• Robots can be imported from URDF or MJCF, and scenes from CAD or USD
• The physics engine can be either PhysX or Newton
• Sensors such as cameras, depth, RTX LiDAR, and radar can be simulated
• Integrates with ROS 2 and can generate synthetic data with Replicator

For reinforcement learning and imitation learning, we used Isaac Lab in combination with Isaac Sim.

GEAR SONIC (L3: Reproducing Ideal Motion Without Falling)

What Kimodo generates is, after all, an ideal motion. Sending this directly to a robot will not produce stable behavior. It will almost certainly fall over.

This is where GEAR SONIC comes in.
It controls all the joints of the body to maintain the center of gravity and ground contact while tracking the target motion.

• Input: Human pose (SMPL), G1 reference motion, teleoperation via VR headset (PICO)
• Internal representation: Inputs are aligned into 64-dimensional latent tokens, and the whole body is controlled at 50 Hz. This is why different types of input can be handled by a single model.
• Training: Motion tracking is formulated as a reinforcement learning task and trained with PPO.
• Deployment to hardware: Inference runs in C++ with TensorRT, and is also compatible with Jetson AGX Orin.

World Models (Not Used This Time)

I also touched on world models, which have been attracting attention recently.
Let me be upfront: they were not used in this proof of concept. They are introduced as a technology to try next.

The use cases are generating synthetic data and making simulated video look closer to real footage. I believe this is a technology that addresses the fundamental physical AI challenge of "insufficient data."

Relationships Between Technologies

Here I summarize the technologies covered so far into a single flow.

1. Give instructions in words: Write the desired action as a sentence.
2. Generate motion with Kimodo: Output whole-body motion using the G1 skeleton.
3. Test in Omniverse: Verify on the digital twin for falls and collisions.
4. Whole-body control with SONIC: Follow the motion while maintaining balance.
5. Run on the actual G1 hardware.

If it doesn't work in virtual space, go back to step 3.
Fix it in virtual space before breaking the actual hardware.
We went through this loop many times.

What We Learned by Trying It on Site

Finally, I talked about what we learned from the proof of concept conducted in an actual building.
This was the part I most wanted attendees to take away from the session.
What I felt more than the performance of the models were the on-site conditions—things you simply cannot know without trying them in the field.

The Gap Between Simulation and Real Hardware Remains

Floor slippage, the weight of objects, joint heat.
Even for motions that succeeded in simulation, fine-tuning work remains when running on actual hardware.
Virtual space reduces the number of failures, but it does not eliminate the need for adjustment on real hardware.

General-Purpose Motion Is Still Difficult

Movements such as walking and crouching can be handled by foundation models.
On the other hand, reliably grasping heavy objects still required considerable individual adjustment.
The stage where a single model can do everything still seems to need a bit more verification.

Summary

Here is what I wanted to convey in the session.

• Physical AI is premised on "failing in virtual space first"
• Test with Omniverse / Isaac Sim, generate motion with Kimodo, and bring it down to the body with SONIC. Next, use world models to compensate for the lack of on-site data.
• The real hurdles to deployment in the field are power supply, communications, and safety design.

Classmethod joined the NVIDIA Cosmos Coalition in July 2026, and in September was recognized as a Solution Advisor in the NVIDIA Partner Network (NPN).
We support everything from building the foundations of robot learning to implementing and operating autonomous behavior in the field.
If you are considering leveraging physical AI, please feel free to reach out.

References

• Classmethod Begins Joint Research on Physical AI for Fire and Disaster Prevention with Yamato Protec
• Classmethod Joins the NVIDIA Cosmos Coalition
• Kimodo (GitHub: nv-tlabs/kimodo)
• Kimodo (NVIDIA Research Project Page)
• GEAR SONIC (GitHub: NVlabs/GR00T-WholeBodyControl)
• SONIC Paper (arXiv:2511.07820)
• NVIDIA Isaac Sim (GitHub)
• NVIDIA Isaac Sim Documentation
• Isaac Lab Documentation
• NVIDIA Omniverse Developer Site
• NVIDIA Cosmos (GitHub)
• Fire and Disaster Management Agency "Fire Fighting Technology Strategy Vision" Overview (PDF)
• Classmethod, NVIDIA Partner Network Partner Page