Observe
The pilot reads the relative positions of a red square and a blue circle in simple synthetic top-down diagrams.
GIEMON · 儀右衛門 / PHYSICAL AI RESEARCH
Intelligence for the mechanisms of the world.
Giemon is AWAI’s model research into spatial understanding and reasoning about action: what is in an image, where it is, and how direction connects observation to movement.
Giemon Spatial 2B v0.1 · Experimental / API not released
Built on NVIDIA Cosmos
THE NAME / KARAKURI GIEMON
The idea behind Giemon.
The name draws inspiration from inventor Hisashige Tanaka, known as Karakuri Giemon. His writing automata and myriad-year clock embody a craft of combining intricate mechanisms to bring motion to life.
Our research direction is intelligence that reads spatial relationships, reasons about mechanisms, and predicts the effects of an action. Beginning with small spatial tasks, we aim to extend evaluation toward tool use, assembly, and simulation.
About Hisashige Tanaka, the inspiration for the name (Japanese) →A first step in reading space.
Structure
Japanese and English instructions produce JSON describing the relative position and image-plane direction toward the other object.
Research toward action
Future evaluation will explore occlusion, change over time and action-conditioned outcomes. Current outputs are not robot control commands.
Measured on a deliberately small task.
Fine-tuned on 64 synthetic diagrams and evaluated after reloading on 16 held-out layouts. Initial experiment: 8 September 2026.
- Content accuracy (against the same converted base)
- 9 / 16 → 16 / 16
- Exact match including JSON format
- 0 / 16 → 16 / 16
- Blank-image control
- 4 / 16
Formatting and content are scored separately. This is a small same-generator test, not a real-world capability measure or a general benchmark score.
Model details and current limits.
- Adaptation
- LoRA · Spatial 2B v0.1
- Availability
- Research pilot. No public model API or weight distribution yet.
The Mac conversion includes a correction for inconsistent parent configuration and weights. Numerical parity with the official runtime is unverified. Real photographs, video, real robots and video-generating world-model capability have not been validated.
Build grounded evaluations with us.
We welcome collaboration on robotics, simulation and visual-understanding evaluation.
Contact AWAI