Google DeepMind Develops Advanced Systems to Improve Robotic Dexterity | Festina Lente - Your leading source of AI news | Turtles AI

Google DeepMind’s recent developments include two AI systems, ALOHA Unleashed and DemoStart, which aim to improve the ability of robots to perform complex tasks through precise and coordinated movements. These advances reflect a commitment to making robots more adept at handling objects in real-world environments, leveraging simulations and learning from demonstrations.

Key Points:

• ALOHA Unleashed improves dual-arm robotic manipulation. 
• DemoStart streamlines simulation learning with fewer demonstrations. 
• Teleoperation and simulations accelerate robotic training. 
• Google DeepMind integrates reinforcement learning with demonstration-based approaches.

Robotic dexterity is a major challenge in robotics, despite many of the complex tasks humans perform on a daily basis being automatic and almost instinctive. Tasks like tying a shoelace or tightening a screw require weeks of intensive training for a robot to perform successfully. Google DeepMind has been working to address this issue by developing two systems that aim to improve the agility of robots: ALOHA Unleashed and DemoStart.

The ALOHA Unleashed system is designed to improve the capabilities of robots in dual-arm manipulation, an area where most current robotic systems are limited. While single-arm manipulation of objects has already been largely developed, the coordinated use of two robotic arms is a step towards the ability to perform more complex tasks. Google DeepMind has said that, thanks to this system, its robots have been able to perform sophisticated tasks such as tying a shoelace, hanging up a shirt, or even repairing other robots. ALOHA Unleashed builds on the evolution of the ALOHA 2 platform, which in turn originates from the open source project developed at Stanford University for bimanual teleoperation. This new version allows robots to learn with a reduced number of demonstrations thanks to the teleoperation of robotic hands, allowing for more efficient training. Among the system’s innovations, there is also the use of a technique inspired by diffusion models, which predicts the robot’s actions starting from random noise, similar to how image-generating AI models work. In addition to improving the robot’s learning, this approach allows to reduce the need for prolonged physical interactions.

In parallel, the DemoStart system represents an advanced solution for simulation-based robotic training. One of the major challenges in teaching agile movements to a multi-fingered robotic hand is the precise control of each joint and sensor. DemoStart addresses this complexity by using a reinforcement learning algorithm that can simulate complex behaviors that can then be transferred to the real world. DemoStart is unique in its ability to learn from simple states, progressively increasing complexity as the system becomes more proficient. This approach reduces the number of demonstrations needed, cutting training times by 100% compared to traditional methods. DemoStart has shown excellent results in simulations, achieving a success rate of over 98% in tasks such as reorienting colored cubes or tightening nuts and bolts. In the physical world, the system also demonstrated high performance, with a 97% success rate in lifting and reorienting cubes and a 64% success rate in a precision task such as inserting a plug into a socket.

DemoStart was developed using the open-source physics simulator MuJuCo, and its approach helps reduce the gap between the simulated and real worlds through domain randomization techniques. This means that once trained in simulation, robots can apply what they have learned almost immediately to the physical world, with knowledge transfer without the need for further training. A major advantage of this method is the potential for reduced development costs and time, as virtual training eliminates many of the practical difficulties associated with physical experiments. However, creating realistic simulations that effectively translate to the real world remains a significant technical challenge.

To demonstrate the system’s capabilities, Google tested DemoStart on a three-fingered robotic hand, called DEX-EE, developed in collaboration with Shadow Robot. Although human-level dexterity is still far away, the progress made so far represents a significant step forward.

With these developments, Google DeepMind continues to explore ways to improve robotic learning, combining reinforcement learning techniques with the use of simulations, making robots more adept at performing complex and dynamic tasks.