OS3The robot & the stack |
Version | v0.2 | |
|---|---|---|---|
| Updated | |||
| Company | OS3 Inc. | Lab | the drop |
Hardware Choices: A mobile robot with two arms, designed in-house.
.----.
| oo |
'----'
||
_____||_____
| || |
_| || |_
| | || | |
| | || | |
|_| || |_|
||
___||___
| |
| O O |
'--------'
The Flywheel: Controlling the hardware to standarize the data. Then training the model to compound it.
┌──────────┐ ┌──────────┐ ┌────────────┐
│ OUR │ ──► │ DEPLOYED │ ──► │ REAL WORLD │
│ HARDWARE │ │ ON SITE │ │ DATA │
└────┬─────┘ └──────────┘ └─────┬──────┘
▲ │
│ ┌──────────┐ ▼
└────────────┤ OUR │◄─────────┘
│ MODELS │
└──────────┘
| Spec | Detail |
|---|---|
| Runtime | 9 hours battery life, extendable to 18 |
| Manipulation | Two arms, ~5 kg per arm, a vertical lift spanning ground to 6 ft |
| Sensing | Camera & LiDAR |
| Safety | Collision avoidance, auto power cutoff, emergency stop |
| Build | Designed & built in the US |
| Everything else | under the cover |
Our research thesis: a model should first learn from the data the world already has, like internet video and first-person footage of people doing ordinary tasks with their hands. By the time it reaches a floor, it already understands the work. From there, minimal robot data and reinforcement learning do the rest.