PROMETHEUS
The Robot's Perfect Pick
Say "pick the blue wrench." The robot finds it, understands its 3D shape, computes a grasp, and picks it up — first try, on an object it has never seen. 85% success on novel objects.
TOTAL ADDRESSABLE MARKET
$6.75B
TIER 1 // 03
PICKING IS STILL DUMB
Modern warehouses handle 100K+ SKUs. Every new object requires hours of per-object training — custom grasps, hand-tuned force profiles, brittle pose estimators. The industry standard sits at 65% pick success. One in three attempts fails, jams the line, and costs real money.
- 100K+
- SKUS PER WAREHOUSE
- $50K
- TRAINING COST PER TASK
- 65%
- INDUSTRY PICK RATE
LANGUAGE IN. GRASP OUT.
PROMETHEUS chains five models into one zero-shot grasping pipeline. A language command goes in. A physically executed pick comes out. No per-object training. No CAD models. No manual pose annotation. The robot sees, understands, and grasps — first try.
MODULE CHAIN:Language Command → AZOTH → PROTEUS → PANOPTES → ERGON → CHIRON → Grasp
- 01
AZOTH
Open-Vocabulary Detection
Takes a natural language command and locates the target object in the scene. "Blue wrench" or "third bottle from the left" — any description, any object.
- 02
PROTEUS
Pixel-Perfect Segmentation
SAM 2.1 produces a sub-pixel instance mask of the target. The robot knows exactly where the object ends and the background begins.
- 03
PANOPTES
Metric Depth Estimation
Converts the 2D scene into metric 3D geometry. Real-world distances, surface normals, and spatial relationships — from a single RGB frame.
- 04
ERGON
6-DoF Grasp Pose (CVPR 2024)
Generates collision-free 6-DoF grasp candidates ranked by stability. Published at CVPR 2024. Works on novel objects with zero training.
- 05
CHIRON
1 kHz Force Control
Closes the loop at 1 kHz with adaptive impedance control. Handles fragile glass and heavy steel with the same policy. Grip force converges in under 50 ms.
FOR LOGISTICS & MANUFACTURING
Every warehouse, factory, and fulfillment center runs on picking. Today, deploying a new SKU means days of integration work. PROMETHEUS makes pick-rate improvements immediate — drop in a new object and the robot handles it on sight. No retraining. No downtime.
- 01Zero per-SKU training eliminates the integration bottleneck
- 0285% pick success where the industry averages 65% — fewer jams, higher throughput
- 03Language-driven commands let operators retask robots in seconds, not days
KEY NUMBERS
- 85%
- NOVEL OBJECT SUCCESS
- 1kHz
- FORCE CONTROL LOOP
- 0TRAINING
- PER SKU REQUIRED
- $6.75B
- ADDRESSABLE MARKET
WHERE THE VALUE IS
| WAREHOUSE BIN PICKING | $3.75B |
|---|---|
| KITTING & ASSEMBLY | $2B |
| RETAIL SHELF STOCKING | $0.5B |
| FOOD & PHARMA HANDLING | $0.5B |
THE MOAT
Language + geometry + control in one pipeline. The industry treats picking as a vision problem. We treat it as a five-model chain — from open-vocabulary detection through metric depth, 6-DoF pose, and 1 kHz force control.
85% where everyone else gets 65%.