bhaprayan

Thanks! Glad to hear it was fun and that you loved that line. Thought it would be the simplest way to describe the underlying concept. Did you have any elements / challenges in mind that may be fun?

I was primarily thinking about more sources, obstacles, and sinks. Right now the solver picks a random solution so it may be interesting to optionally make that deterministic too.

Understanding intent is really hard indeed. What’s helpful about constraining it to you showing the effect is if you are able to do it, there exists at least one possible mechanism.

Thanks, happy to hear you liked it!

Really neat demo! I skimmed through the paper and it looks like they’re using a precomputed database to find an assembly that matches the motion (presumably as a seed to instantiate the solver). That’s definitely more efficient than what I have right now (though perhaps less versatile). I like that they manufactured a few of the characters using 3D printing! That was definitely on my mind, really happy to see they did it.

Some of the mechanisms (belt drives etc) might be a little tricky with 3D printers, but would be very cool to see the mechanism tangibly realized.

Happy to open source, the code is a little messy right now, so I’ll clean it up and post a link in a few days!

High level:

I’m using ThreeJS for the rendering and Rapier (https://rapier.rs/) for simulating bodies, joints, collisions, forces, etc. The mechanism is solved for using a simple heuristic search (eg constraints, etc that you introduce in the scene) and then run through Rapier to find a configuration that works, though there are definitely more sophisticated ways to do this. The audio is based off a sound pack which I tried to make physically realistic so impacts are panned relative to the camera, with compression and a small generated reverb.

Haha it sort of came out before I was born, otherwise I definitely would have! I think I'll take a leaf out of their book and expose a mode to make the simulation deterministic. I'm using a PRNG right now to make the mechanism selection more "creative", but I could see determinism being an interesting experience.

Thanks, appreciate it!

Embarrassed I'm just finding out about this, very cool! Neat that they simulated air pressure too. I should try that, would be fun to influence the texture of the sound.

That was cool! Watched the demo video, didn't go through the paper yet. I'm running a simple heuristic to test for candidate solutions and running them through Rapier (https://rapier.rs), but simulated annealing to efficiently search the configuration space would be a nice extension.

Thanks! Great idea, I hadn’t realized the interaction is binary. It would be cool to allow for "continuous" inputs too. The world would have to account for interactions over a recording period right.

Behind the scenes, I’ve tried to make the audio production realistic, so there’s some acoustic modeling going on (eg echoes, attenuation of sound, etc). But yes you could imagine having the input forces be continuous and having the mechanism trying to figure out how to reproduce the sound exactly.

That's a great point. I suspect though, the critique applies to decision making by consensus more broadly.

Decisions made by consensus are not necessarily the optimal strategy (as defined by efficacy to achieve targets). This method simply reflects that.

However, the benefit of this method is that all participants are forced to be more rigorous about why they hold certain beliefs, and then making an attempt to aggregate those preferences to drive an outcome.

Instantiation of ideas from the paper here (NOVID app): https://www.novid.org/

I attended a talk about this recently, and I honestly think it's a revolutionary / game changing way to do contact tracing.

TLDR is to understand the spread of COVID-19 around you (through second-order network connections and beyond). This makes it much more similar to a weather forecast (e.g. the virus is approaching your vicinity through people in your network). This makes it a much more effective way to systematically identify, isolate, and contain further spread.

The idea is so simple and effective, I can't believe this isn't front and center over the news as a way to effectively contain the spread of COVID (beyond the existing measures of wearing masks, social distancing, etc).

This is amazing! Really excited to see where this goes, good luck :)

This post was inspired by a question Richard Hamming asks in "You and your research" (paraphrasing): "What are the most important problems in your field?". When I thought hard about this question, I realized that I didn’t have a satisfying answer. This post is an ongoing attempt to remedy this.

Since solving these challenges is a collective pursuit, I’d appreciate suggestions on content to add, general feedback, and comments! :)

Short post on using computational graphs to derive update rules for backpropagation, illustrated using a Gated Recurrent Unit (GRU). Questions, feedback, corrections? Do reach out! :)