Disney Research Makes Dynamic Robots Less Wiggly, More Lifelike | Hacker News Reader

Disney Research Makes Dynamic Robots Less Wiggly, More Lifelike

spectrum.ieee.org

93 pointssohkamyung19 comments

Duplicate of https://news.ycombinator.com/item?id=20670932

Disney Research Youtube channel is an amazing place to hang out: https://www.youtube.com/user/DisneyResearchHub/videos

That's great, thanks. Boston Dynamics is a good one too, although less frequent: https://www.youtube.com/user/BostonDynamics

Anyone know of other channels like these?

University of Michigan has an interesting channel: https://www.youtube.com/user/DynamicLegLocomotion

Also Nvidia's YouTube channel occasionally has interesting videos on their Jetson platform.

I wonder if this is a tractable reinforcement learning problem. The objective is somewhat clearly defined-minimize spatial deviation over time from the input path.

In order to train the model, wouldn’t you want a physics simulation, so you can train the model quickly? But if you had a physics simulation, aren’t there much easier methods to use than machine learning?

Who said there was any machine learning?

The comment I was replying to was talking about machine learning.

TLDR:

Active damping

Except if you read the article you'd realise this is not active damping. They're optimising the movements using a physical model so that they don't need active damping at all.

What they are saying in the article is that they are optimizing the movements using a physical model so that they don't need passive damping.. but isn't that exactly what active damping is?

it's preparing the movements in advance via simulation, is it still active damping if it has done that?

They actually mention active damping in the video so yes. It’s less resilient as ware will make the robots inconsistent over time, but the tradeoff for a less costly robot with fewer sensors may be worth it.

I don't think so, no. It would need to be reactive to active damping, and as the article says, there aren't any onboard sensors, so if you push them they'll wobble like crazy.

It's a tricky distinction because the program is changed to dampen the vibrations but it's all baked into the program at start it seems. I'd say it's not active because it doesn't dynamically monitor and change the program. In theory you could perturb these animatronics in such a way that their damping motions actually increase the vibrations.

I think of the term optimal control [1] given that problem description. Maybe there's some overlap in the definitions?

[1] https://en.wikipedia.org/wiki/Optimal_control

It's not active damping, it's predictive, pre-computed powered damping.

If you change the physical behavior of the robot, such as by heating or cooling its arm rods, or by changing the mass of its end-weights, the pre-computed damping movements will be less accurate, and the wobble may return.

Active damping would require sensors and a more complicated microcontroller on the robot, rather than just looped playback from a ROM. Cheaper smooth-moving robots mean more animatronic elements can be put in your feature within a given budget.

Yea you definitely DR

You’re getting downvoted but you’re right!

Reply on news.ycombinator.com