Training humanoid robots is a costly and time consuming endeavour. Imitation learning, the current state-of-the-art method, requires humans to demonstrate tasks over and over before a robot has a chance of successfully recreating them.

​

Tesla pays individuals $48 an hour (close to $100k per year) to show its robot how to act like a human, performing a variety of everyday tasks one after another in a controlled environment.

​

According to Google DeepMind’s Danny Driess, it took 8,000 demonstrations to teach the company’s robot hand how to tie shoelaces and a dozen players over four weeks to teach its ping pong robot to play at an intermediate level.

​

The high labour costs and sheer number of demonstrations associated with teaching a robot every possible innate and learned human behaviour suggests that progress towards general-purpose humanoids will continue to be bottlenecked.

 

Regardless of hardware and software advancements, this fundamentally unscalable data collection model needs reform for humanoid robotics to take the next step.

​

This is Our Mission