A robot that first imagines what will happen before it moves toppled and dropped fewer items in a lab test, Japanese telecoms group NTT says.
NTT announced on 11 November a “world model” that takes several possible actions suggested by a robot’s existing control software, generates a short predicted video of each, and scores how likely each one is to end in a collapsed stack or a dropped object. The robot then carries out the action most likely to succeed. NTT says its addition, which it calls a physics control adapter, tracks each object’s position, orientation, speed and mass and predicts changes using equations of motion, rather than relying on how the video looks alone.
In its own testing, a small robot stacking items on a shelf succeeded 92.3% of the time with the technology, up from 73.1% without it. NTT has not named the robot or published a paper, and the result comes from a test task, not a working site. Its first target is loading and unloading airport baggage of varying shapes and weights, with field trials planned with Japanese robotics company Mujin. No dates were given.
Why it matters: handling mixed loads without dropping or crushing them is one of the harder jobs for robots in warehouses and airports. NTT says the method also copes with situations missing from its training data, which matters where collecting examples is hard. New to the idea? Read What is a world model?
