Teaching robots to navigate in new environments is difficult. You can train them on physical, real-world data from human-made recordings, but that is scarce and expensive to collect. Digital simulations are a fast, scalable way to teach them new things, but the robots often fail when taken from virtual worlds and asked to perform the same tasks in the real world.
Now there may be a better option: a new system that uses generative AI models in combination with a physics simulator to develop virtual training grounds that more accurately reflect the physical world. Robots trained with this method operated with a higher success rate than robots trained with more traditional techniques during real-world testing.
Researchers used the system, called LucidSim, to train a robot dog in parkour, which involved clambering over a box and climbing stairs, despite never seeing any real-world data. The approach shows how helpful generative AI can be when it comes to teaching robots to perform challenging tasks. It also raises the possibility that we could eventually train them in completely virtual worlds. Read the full story.
—Rhiannon Williams
The African AI researchers are ready to start
When we talk about the global race for AI dominance, the conversation often turns to tensions between the US and China, and European efforts to regulate the technology. But it’s high time we talk about another player: Africa.
African AI researchers are forging their own path and developing tools that meet the needs of Africans in their own language. Their story is not just one of perseverance and innovation, but of preserving cultures and fighting to shape the way AI technologies are used on their own continent. However, they face many barriers. Read the full story.
—Melissa Heikkilä