Our latest progress in the dexterity of the robot

Two new AI, Aloha Unleashed and Demostart systems, help robots learn to perform complex tasks requiring skillful movement

People perform many tasks every day, such as binding laces or tightening the screw. But in the case of robots, learning these highly discovered tasks is extremely difficult to do. To make robots more useful in people's lives, they must be better to make contact with physical objects in dynamic environments.

Today we are introducing two new articles containing our latest progress of artificial intelligence (AI) in robot research: Aloha released which helps the robots perform complex and innovative manipulative tasks of two -arms; AND Demostart which uses simulations to improve real performance on many robotic hands.

By helping robots learn from interpersonal demonstrations and explain paintings into action, these systems pave the path to robots that can perform many useful tasks.

Improving the teaching of imitation with two robotic arms

Until now, most of the advanced AI robots were able to receive and place objects with one arm. IN Our new articleWe present Aloha Unleashed, which reaches a high level of dexterity in the manipulation of Bi-Ramnia. Thanks to this new method, our robot learned to tie a lace, hang a shirt, fix another job, insert equipment, and even clean the kitchen.

The Unleashed Aloha method is based on ours Aloha 2 Original platform Aloha (cheap Open Source hardware system for bimanual teleoperation) Stanford University.

Aloha 2 is much more skillful than previous systems, because it has two hands that can be easily telecathized for training purposes and data collection, and allows robots to learn to perform new tasks with fewer demonstrations.

We also improved the ergonomics of the equipment of the equipment and improved the learning process in our latest system. First, we collected demonstration data, remotely supporting the behavior of the robot, performing difficult tasks, such as binding laces and hanging T -shirts. Then we used the diffusion method, predicting the actions of the robot from a random noise, just like our Imagen model generates images. This helps the robot learn based on data, so he can perform the same tasks on its own.

Learning robotic behavior from several simulated demonstrations

Controlling a skillful, robotic hand is a complex task that becomes even more complex with each additional finger, joint and sensor. In another new paperWe present a demostart that uses a reinforcement algorithm to help robots get skillful behavior in simulation. These learned behaviors are particularly useful in the case of complex examples of performance, such as multi -palm hands.

Demostart first learns from easy states and sometimes begins to learn from more difficult states, until he mastered the tasks you can. It requires 100 x less simulated demonstrations to learn to solve the task in the simulation than usual when learning from examples in the real world for the same purpose.

The robot achieved a success rate over 98% for many different tasks in the simulation, including reorientation of cubes with a certain color showing, tightening the nut and screw and tools for cultivation. In the real configuration, 97% of the success rate in reorientation and lifting of the ankle and 64% with the task of inserting a plugin, which required coordination and precision in high smoking.

We developed a demostart from MujocoOur Open Source physics simulator. After mastering a number of tasks in the simulation and the use of standard techniques to reduce the Sim-Real gap, such as domain randomization, our approach was able to transfer almost zero shot to the physical world.

Robotic learning in a simulation may reduce the costs and the time needed to conduct real, physical experiments. But it is difficult to design these simulations, and in addition they do not always translate back back into real results. By combining learning to learn from several demonstrations, the progressive learning of the demostart automatically generates a curriculum that combines the SIM-Real gap, facilitating the transfer of knowledge from the simulation to physical robot and reducing the costs and time needed to conduct physical experiments.

To enable more advanced robot learning through intensive experiments, we tested this new approach on a three hand of work Dex-EEwhich was developed in cooperation with Shadow robot.

The future of the robot's dexterity

Robotics is a unique AI research area that shows how well our approaches work in the real world. For example, a large language model can say how to tighten the screw or tie shoes, but even if it was incorporated into the robot, he would not be able to perform these tasks.

One day AI robots will help people with all kinds of tasks at home, at the workplace and more. Studies of dexterity, including an effective and general approach to science that we described today, will help to make this future.

We still have a long way before the robots are able to capture and operate the objects with ease and precision of people, but we are doing significant progress, and every breakthrough innovation is another step in the right direction.