Related Posts

Comments (21) on “This AI Learned To Animate Humanoids!🚶”

  3. I am trying to help someone who broke their neck in July at C4 (completely paralyzed but breathing on his own). My own brother died many yeas ago after living 18 months with a C2 break (completely paralyzed on a respirator).

    There are many technologies for picking up intentional and automatic signals from the brain, or skin near muscles. But my question is about AI control training and optimization. I think it is much easier to take intention signals and map them to robotics, than to try to map out the existing muscles and autonomic controls.

    Have you looked at this topic at all? For everyday low cost methods for the roughly 300,000 people in the US living with spinal cord injury, just feeding, picking things up with robotic hands, taking medications, adjusting chairs and controlling wheel chairs and cars are the usual "wanted list".

    But the rising number of elderly, many other causes of nerve damage or weakness or paralysis – this raises the number much larger. Ask "how many people in the us living with paralysis" and you get about 5 million. The US is 332 Million, the world is 7900 million. But it does not scale exactly. The problem being that many places in the world, these things lead to death, not permanent disabilities of various sorts, requiring full time or part time care. I can tell you how many people, what they face, what it costs, who is working on it, what has been tried. But this piece of "control the existing muscles as well or better than they were controlled and maintained before" has not been done well. Much of the reason for that, is that groups and individuals all work on their own, making things that benefit them, so most global issues never get solved. If you want, I can tell you about a few thousand such. But now I just want to see if, by summer 2022, if some specifics and examples can be demonstrated. The framework is important – who, why, where – and all the pieces need to fit together without all the parts fighting each other. So it needs to be open, lossless, auditable.

    Ignore where you get the signals, it is possible to 3D scan anyone and "rig" the 3D model for animation. It is also possible to map the muscles, tendons and details of the joints and activation involved. Yes, I know, it is tedious and time consuming. But a few months of scanning would be tiny compared to the lives some live. So can you think how to create a global community to solve this once and for all? With low cost sustainable methods. You must know "constrained optimization" in many forms. Modeled processes that you solve for one or several objectives. So imagine solving the global economic, social, financial, organizational and political issues all at once, and one part is helping these kinds of individuals to live with dignity and purpose.

    Right now most of this kind of retraining – trying to use what little connection is left, is being handled by human rehabilitation people and volunteers. But all those neural (both automatic and intentional) pickups and stimulation of current muscles are being covered by piecemeal methods globally. The brain machine interfaces, most of the money is going into surgery and expensive and intrusive methods. But there are about 40 basic methods for 3D imaging of nerve activity that can give varying levels of data.

    The groups now are happy if they can bang bang the nerves and help people to stand, and take a few clumsy steps. But there is enough data to control the original muscles, or by-pass and do the equivalent or better machine controlled devices and processes. A non-invasive neural interface for machine welding would have the sensors and speed of a robot welder, but the control and experience of a human. Yes, you could use learning methods to replace the humans, but humans are really inexpensive now. That is why Zooniverse can get millions of volunteers to do human-in-the-loop recognition problems that a decent algorithm could handle.

    Richard Collins, Director, The Internet Foundation

  5. I wonder if you could record a bunch of motion capture data of an actor doing a bunch of random things, then input that into the motion AI so that every character has their own body language that really matches their character properly, while smoothly moving around a complex environment!

  15. Oh man, that hind-leg-sliding, I see that a lot in the game Planet Zoo. Drives me nuts.
    (I'm a gamer by choice and just started watching your fluid-simulation videos. Cannot wait for the next 10 years of computing evolution.

Comments are closed.