New DeepMind AI model can control robotic actions it was never trained to do

You are currently viewing New DeepMind AI model can control robotic actions it was never trained to do
<span class="bsf-rt-reading-time"><span class="bsf-rt-display-label" prefix=""></span> <span class="bsf-rt-display-time" reading_time="2"></span> <span class="bsf-rt-display-postfix" postfix="min read"></span></span><!-- .bsf-rt-reading-time -->

Google DeepMind has a new AI model that can direct robotic tasks it was never trained to perform.

Named RT-2, the model learns from web and robotics data. It then turns this information into simple instructions for machines.

In tests, the model was asked to take actions never seen in the robotic data, such as placing oranges in a matching bowl. To follow these commands, the system had to translate knowledge from web-based data. According to DeepMind, the model had a 62% success for these operations — double that of its predecessor, RT-1.

“Just like language models are trained on text from the web to learn general ideas and concepts, RT-2 transfers knowledge from web data to inform robot behaviour,” said Vincent Vanhoucke, head of robotics at DeepMind. “In other words, RT-2 can speak robot.”

The model was tested on various emergent robotic skills that are not present in the robotics data and require knowledge transfer from web pre-training
RT-2 was tested on various robotic skills that weren’t present in the robotics data. Credit: Google DeepMind

The tests showed RT-2 has impressive generalisation capabilities. It also has an improved semantic and visual understanding of robotic data that wasn’t previously encountered.

Notably, the model can use rudimentary reasoning to follow new user commands. Impressively, it can even perform multi-stage semantic reasoning. For instance, when instructed to pick an object that could be used as a hammer, RT-2 correctly identified a rock as the best option.

Here we show an example of such reasoning and the robot’s resulting behaviour
In one test, RT-2 figured out that a rock would be the best object to pick up as an improvised hammer. Credit: Google DeepMind

In another evaluation, the model was commanded to push a bottle of ketchup towards a blue cube.

There were several items in the scene, but the only one in the training dataset was the cube. Nonetheless, RT-2 successfully pushed the ketchup towards the specified destination.

RT-2 performs well on real robot Language Table tasks. None of the objects except the blue cube were present in the training data.
RT-2 performed well in real-world tasks. Credit: Google DeepMind

DeepMind has heralded RT-2 as a breakthrough in artificial intelligence. The Londonlab says the model brings us closer to a future of helpful robots.

“Not only does RT-2 show how advances in AI are cascading rapidly into robotics, it shows enormous promise for more general-purpose robots,” said Vanhoucke. “While there is still a tremendous amount of work to be done to enable helpful robots in human-centered environments, RT-2 shows us an exciting future for robotics just within grasp.”

You can read the RT-2 study paper here.


Back to top