|
Briefing
|
- Dactyl – Non-profit OpenAI developed new robotic hand system, called Dactyl, that learned to grasp block like human through trial and error from simulation and then in real world, without human input
- Reinforcement Learning – Rewarded robot when it made right actions, such as rotating block, and punished when it fails goal, such as dropping object
- Learning to See – Armed with three cameras, Dactyl can see by estimating position and orientation of object on hand
- Object Manipulation – Learned how to pivot block through fingers, sliding block through palm, and rotating block with fingers
- Relevance – Innovation one step closer to making general artificial intelligence where robot hand can grasp new objects without requiring new data, similar to humans
|
