ATHENA
Learn From One Demo
Show a robot a task once. One human demonstration. 87.8% accuracy. Conventional methods need 100 demonstrations to reach 57.2%. That's 100× data efficiency.
ONE-SHOT SUCCESS RATE
87.8%
TIER 2 // 04
ROBOT PROGRAMMING IS STUCK
Every new task a robot learns costs $5K–$50K in integrator time. Enterprises are locked into system integrator dependency — months of lead time, rigid programs that break when a single object changes. Conventional imitation learning needs 100+ demonstrations per task and still only reaches 57.2% accuracy. The economics don't scale.
- $50K
- COST PER NEW TASK
- 100
- DEMOS NEEDED TODAY
- 57.2%
- CONVENTIONAL ACCURACY
SHOW ONCE. DEPLOY FOREVER.
ATHENA uses a two-phase approach. Phase 1 captures a single human demonstration — Gaussian splatting the scene, segmenting objects, labeling with language, tracing the hand motion, and estimating 6DoF pose. Phase 2 generalizes: a 450M-parameter Vision-Language-Action model maps perception to motion, while a 1kHz compliant controller keeps the robot safe in the real world.
MODULE CHAIN:DEMO: GENESIS → MONAD → AZOTH → DAEMON → ERGON // EXECUTE: PYGMALION → CHIRON
- 01PHASE 1 — DEMO CAPTURE
GENESIS
3D Gaussian Splat
Reconstructs the demonstration scene as a photorealistic 3D Gaussian field — a transferable spatial memory of how the task was performed.
- 02PHASE 1 — DEMO CAPTURE
MONAD
Segment & Track
Isolates every object in the scene with persistent masks. The robot knows what moved, what was grasped, and what stayed still.
- 03PHASE 1 — DEMO CAPTURE
AZOTH
Language Labels
Grounds every segment to natural language. Not "mask #12" but "blue hex bolt, M6, left tray." Enables verbal task specification.
- 04PHASE 1 — DEMO CAPTURE
DAEMON
Motion Trace
Extracts the full 3D hand trajectory from the demonstration video — the motion skeleton the robot will replicate.
- 05PHASE 1 — DEMO CAPTURE
ERGON
6DoF Pose Estimation
Computes precise 6-degree-of-freedom pose for every manipulated object. Orientation, position, approach angle — all captured.
- 06PHASE 2 — EXECUTE & GENERALIZE
PYGMALION
450M VLA Policy
A 450-million-parameter Vision-Language-Action model that maps scene perception + language instruction to motor actions. The brain that generalizes.
- 07PHASE 2 — EXECUTE & GENERALIZE
CHIRON
1kHz Compliance
Real-time compliant controller running at 1kHz. Absorbs contact forces, adapts to object variation, keeps the robot safe during execution.
FOR ROBOTICS
This is the iPhone moment for robot programming. Instead of hiring an integrator for $50K and waiting three months, a floor manager shows the robot the task once. Five minutes later it's running. New product variant? Show it again. ATHENA turns robot deployment from a capital project into a daily workflow.
- 01One-shot learning: 87.8% accuracy from a single human demonstration
- 02No integrator needed — factory staff teach the robot directly
- 03Edge-deployable: 450M model runs on-device, no cloud dependency
KEY NUMBERS
- 87.8%
- ONE-SHOT ACCURACY
- 100×
- DATA EFFICIENCY
- 5MIN
- TASK ACQUISITION TIME
- $19B
- ADDRESSABLE MARKET
WHERE THE VALUE IS
| HOME ROBOTICS | $11B |
|---|---|
| SMALL-BATCH MANUFACTURING | $5B |
| CARE & ASSISTIVE | $3B |
| FIELD SERVICE & MAINTENANCE | $0.5B |
THE MOAT
One demo. 87.8%. The Gaussian Splat is the transferable knowledge — a 3D memory of how a task was performed that generalizes to new objects and new scenes.
One demo. Not one hundred.