Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

Introducing 2️⃣ new AI systems for robotics: 🤖 ALOHA Unleashed to perform two-armed manipulation tasks 🦾 DemoStart to control a multi-fingered robotic hand They learned to tackle a range of actions requiring dexterity. Here's how. 🧵

297,118 Aufrufe • vor 1 Jahr •via X (Twitter)

10 Kommentare

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

Manipulating objects using just one robot arm is challenging. Enter ALOHA Unleashed, which builds upon our ALOHA 2 system. With two arms, it can be teleoperated to collect high quality training data. With this system, robots can perform new tasks with fewer demonstrations.

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

We improved the hardware’s ergonomics while combining ALOHA 2 with a diffusion method - which generates realistic actions from random noise - to teach complex skills. This enabled our bi-arm system to push the boundaries of learning to accomplish difficult tasks autonomously.

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

To challenge our robot, we set out to see if it could perform in both simulation and the real-world. Our results show it was able to: ▪️Tie shoelaces ▪️Insert gears ▪️Replace parts on another robot ▪️Hang up shirts ▪️Clean up a kitchen

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

An AI controlling a multi-fingered robotic hand means it could carry out more useful, physical actions. DemoStart takes us one step closer by using a reinforcement learning algorithm to master various behaviors from just a handful of simulated demonstrations.

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

Over time, DemoStart considers increasingly difficult states in simulation until it masters a task. This staged learning approach enables it to successfully transfer knowledge to the real world - reducing the cost and time needed for running physical experiments.

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

We tested it on the DEX-EE hand, developed with @shadowrobot. In simulation, the robot had to reorient a 3D cube to show a certain color. It then performed this task in the real world 97% of the time. It was also able to reliably insert a plug into a socket in both set-ups.

Profilbild von Google DeepMind
Google DeepMindvor 1 Jahr

There’s still a long way to go for robotics to achieve human-level dexterity but the future is exciting. Our research could pave the way in creating more helpful, dexterous robots that could one day assist in the home, office and beyond. Find out more →

Profilbild von Kayla Cardillo
Kayla Cardillovor 1 Jahr

Shouldn’t robots have their own way of doing things? Making the robots copy humans just looks slow

Profilbild von Architect🛡️
Architect🛡️vor 1 Jahr

i also hang my shirts this way superintelligence 🤝

Profilbild von kache
kachevor 1 Jahr

thank you for sharing!!

Ähnliche Videos