Robots find it hard to catch objects thrown at them because each object has different physical properties, like how it bounces or deforms. The robot must react quickly using only a single camera, and humanoid robots have weak arms that can break if they move too fast.
To solve this, researchers trained a catching policy in a simulation. They used careful reward design to encourage good grasps and a training curriculum that gradually increased difficulty while keeping the robot within its strength limits. The policy was then transferred directly to a real humanoid robot without any real-world training.
The robot successfully caught nine objects it had never seen before, including soft toys and chip bags, in both underhand and overhand catches. This shows that simulation-based training can generalize to real-world objects, making humanoid robots more capable and safer for dynamic tasks.