Almost all progress at doing tasks reliably has been made with 2 massive caveats:
1. Force. Walking, running, fighting, doing backflips, etc. all allow for large amounts of force, without a lot of dynamic precision required. Many common tasks require precise and dynamic force. E.g. for washing a window, pushing too hard breaks the glass while pushing too softly will leave streaks.
2. Environment interaction. Most reliable humanoid robots do minimal environment interaction beyond self-balancing. They walk/run/jump in environments that are largely open, with usually convex blocky obstacles. The real world has lots of tasks that require processing beyond low-resolution maps of solid/open space. E.g. I'd want to see a robot that can walk through a forest: jumping/stepping over thin branches that are hard to see, ducking under fallen logs, pushing though bendy branches without breaking them, avoiding ground that is muddy, and seeing through the current obstacle to determine if the obstacle beyond is traversable.
Just to reiterate, I don't see fast progress being made on doing these tasks reliably. It's easy to show 1/N success rate, and much much harder to show ~N/N success rate on these dynamic tasks.
1. Force. Walking, running, fighting, doing backflips, etc. all allow for large amounts of force, without a lot of dynamic precision required. Many common tasks require precise and dynamic force. E.g. for washing a window, pushing too hard breaks the glass while pushing too softly will leave streaks.
2. Environment interaction. Most reliable humanoid robots do minimal environment interaction beyond self-balancing. They walk/run/jump in environments that are largely open, with usually convex blocky obstacles. The real world has lots of tasks that require processing beyond low-resolution maps of solid/open space. E.g. I'd want to see a robot that can walk through a forest: jumping/stepping over thin branches that are hard to see, ducking under fallen logs, pushing though bendy branches without breaking them, avoiding ground that is muddy, and seeing through the current obstacle to determine if the obstacle beyond is traversable.
Just to reiterate, I don't see fast progress being made on doing these tasks reliably. It's easy to show 1/N success rate, and much much harder to show ~N/N success rate on these dynamic tasks.