Nvidia’s self-training robots master tricky tasks with AI coding agents
A team of researchers from Nvidia, Carnegie Mellon University, and UC Berkeley has demonstrated a breakthrough in robotics: robots that teach themselves how to perform complex, dexterous tasks using AI coding agents. Instead of relying on pre-programmed instructions, these robots generate their own training routines, refine their skills through trial and error, and achieve success rates as high as 99 percent on demanding manipulation challenges.
A new approach to robotic learning
The method centers on AI coding agents that write and execute Python scripts in real time. These scripts guide the robots’ movements, adjust their strategies based on feedback, and even rewrite their own code to improve performance. By running thousands of trials autonomously—grasping objects, stacking blocks, or handling delicate items—the system iteratively optimizes both the robot’s actions and the underlying control logic.
Scaling autonomy in the physical world
What makes this approach notable is its application in the real world, not just simulations. A fleet of eight physical robots was deployed to test the system, performing tasks that traditionally require precise human-like dexterity. The high success rates suggest that self-improving robotic systems could soon move beyond controlled lab environments into warehouses, factories, and even homes—where adaptability and continuous learning are critical.
The research highlights a shift toward robots that don’t just follow instructions but evolve their own solutions. As AI coding agents become more capable, the boundary between programmer and machine may blur further—ushering in an era where robots design their own training, and ultimately, their own capabilities.
Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.