A pile of laundry sits in the washing machine, waiting to be hung out to dry. Another pile of clean clothes lies haphazardly on the bed. The wet clothes have to be taken out and hung on the washing line and the pile on the bed has to be dealt with by folding each item and keeping them in the wardrobe. And all this is to be done by robots.

This was a real-world challenge in the home scenario competition at the recently concluded 2nd World Humanoid Robot Games in Beijing. The robot participants had to complete the entire sequence autonomously and continuously. No points were awarded if the process was interrupted.

The winning team completed the sequence in 26 minutes, a complex autonomous operation almost unimaginable last year. In industrial settings, robots handle tasks such as transporting materials, stocking shelves, picking items, and packing them for storage. Hotel-based competitions take place in actual hotels, where robots navigate real-world obstacles while performing multi-step tasks such as transporting luggage, restocking rooms and tidying up.

Multi-scenario competitions are continuously expanding the capabilities of humanoid robots, from being remote-controlled manually to independently completing entire task sequences.

If they can successfully complete multiple tasks in the competition arena, how long is it before we have robot butlers? Wang He, founder and CTO of the robot company Galbot, addressed this common question. He said the core issue lies in the robots' insufficient capability of generalization across different robotic applications. "Being able to fold one specific garment doesn't mean a robot can handle another style," he said.

He explained further, "It's not that we lack the ability to (make robots that) perform the tasks. Rather, adapting to each new scenario requires a massive investment of time and R&D resources to collect and process the necessary data, which can actually exceed the value of the task itself."

The key to breaking this deadlock is continuous evolution of foundation models. Currently, completing a task may require two days of data collection. In six months, that could drop to just five hours. Within two years, robots may no longer need to gather prior data, enabling them to learn instantly when they see a task or hear a command.

This evolution will bring a qualitative shift in cost and efficiency. "Only when costs are low enough and efficiency is high enough will a laboratory technology evolve into a product that ordinary people can afford and use with ease," Wang said.