5 ms·

Would it be incorrect to say that the primary problems Boston Dynamics works on, which is planning and controls, is already mostly a solved problem? My underst

Would it be incorrect to say that the primary problems Boston Dynamics works on, which is planning and controls, is already mostly a solved problem? My understanding is that the majority of research was completed within the past few decades, in comparison to the relative nascency of modern deep learning techniques.

rm445 5y ago
The foundations are understood, in a 'chemistry is just the Schrödinger equation' kind of way. Any amount of research and engineering is still needed to make machines which can move themselves effectively, be energy-efficient, and interact with the messy and unstructured wider world.
- lawrenceyan 5y ago
  How far off would you estimate a company like Boston Dynamics is from production then? (Say with existing levels of industrialization in the form of non-humanoid robots already commonly used as a benchmark/baseline)
  - rm445 5y ago
    My guess (as an engineer but not in that field) is that some sort of breakthrough is needed or it will be a long slog. The general problem of robots navigating arbitrary spaces will probably be solved by wider advances in machine learning. Because a robot mapping its environment, planning its motion, and carrying out that motion while adapting to the dynamic environment, isn't being treated as a classic control systems problem any more, but fed into trained nets. It's too broad a problem for even sophisticated control systems, in the general case. Even once robots can do amazing things, there needs to be an economic use case that's compatible with what they can do, how they're powered, and how long they can work for. Similar to quadcopters in a way - they have changed the world in some limited ways, but so far they haven't scaled up to carry people or out to deliver parcels, so they remain niche. If someone figures out one day that robots can deliver burritos/assassinate dissidents/stack shelves efficiently, it could be a quick takeoff.

rm445 5y ago

The foundations are understood, in a 'chemistry is just the Schrödinger equation' kind of way. Any amount of research and engineering is still needed to make machines which can move themselves effectively, be energy-efficient, and interact with the messy and unstructured wider world.

lawrenceyan 5y ago
How far off would you estimate a company like Boston Dynamics is from production then? (Say with existing levels of industrialization in the form of non-humanoid robots already commonly used as a benchmark/baseline)
- rm445 5y ago
  My guess (as an engineer but not in that field) is that some sort of breakthrough is needed or it will be a long slog. The general problem of robots navigating arbitrary spaces will probably be solved by wider advances in machine learning. Because a robot mapping its environment, planning its motion, and carrying out that motion while adapting to the dynamic environment, isn't being treated as a classic control systems problem any more, but fed into trained nets. It's too broad a problem for even sophisticated control systems, in the general case. Even once robots can do amazing things, there needs to be an economic use case that's compatible with what they can do, how they're powered, and how long they can work for. Similar to quadcopters in a way - they have changed the world in some limited ways, but so far they haven't scaled up to carry people or out to deliver parcels, so they remain niche. If someone figures out one day that robots can deliver burritos/assassinate dissidents/stack shelves efficiently, it could be a quick takeoff.