Teaching robots to navigate new environments is difficult. You can train them on real physical data extracted from recordings made by humans, but this is rare and expensive to collect. Digital simulations provide a quick and scalable way to teach them to do new things, but robots often fail when they are removed from virtual worlds and asked to perform the same tasks in the real world.

Now there is potentially a better option: a new system that uses generative AI models in conjunction with a physical simulator to develop virtual training grounds that more closely reflect the physical world. Bots trained using this method performed with a higher success rate than those trained using more traditional techniques in real-world testing.

Researchers used the system, called LucidSim, to train a robot dog in parkour, making it climb a box and run up stairs, even though it never saw real-world data. This approach demonstrates how useful generative AI could be when it comes to teaching robots to perform difficult tasks. It also raises the possibility that we could eventually train them in entirely virtual worlds. Read the full story.

—Rhiannon Williams

African AI researchers are ready to take off

When we talk about the global race for AI dominance, the conversation often focuses on tensions between the United States and China, as well as European efforts to regulate this technology. But it is high time to talk about another actor: Africa.

African AI researchers are charting their own path, developing tools that meet the needs of Africans, in their own languages. Their story is not only one of perseverance and innovation, but also one of preserving cultures and fighting to shape the way AI technologies are used on their own continent. However, they face many obstacles. Read the full story.

—Melissa Heikkilä