Simon Clavet

Our interactive mocap-playing virtual robot can now survive very difficult situations:

@belinghy Yes. Along with @anorangeduck, @_Quazi_, and Felix Harvey, who performed the KungFu routine. To be perfectly clear, we are nerds, not athletes.

My kids identified me instantly

Kungfu BOT: Unitree G1🥳 We have continued to upgrade the Unitree G1's algorithm, enabling it to learn and perform virtually any movement. What other moves would you like to see. Do share with us in the comments. (Please keep a safe distance from the robot.) #Unitree #Kungfu #EmbodiedAI #SpringFestivalGalaRobot #AI #Humanoid #Bipedal #WorldModel #Dance

@siyuanhuang95 The mocap data they used is dance1subject2 of Lafan1 public animation dataset. I was in the suit. Surreal.

The CEO of Unitree, XingXing Wang, posted a dancing video at Rednote against the hype that the previous dance video was AI- or CG- generated. The dance is performed before a mirror and with sound, which makes it 100% real. Really cool demo! #Unitree #Humanoid #RobotDance

Simon Clavet is bringing his talent to TiMi Studio Group! A pioneer in procedural animation for gaming, he now joins our TiMi Montreal team to help us get our unannounced AAA open world multiplayer project in the hands of gamers worldwide. #montreal #gamedev #videogameindustry

@anorangeduck @RH_Hagdahl One thing I tried was to compute pos/vel distance 150ms ahead in current and candidate anims. So we minimise mismatch in the middle of the 300ms blend instead of the start.

@RH_Hagdahl Agreed - probably nice when you are expecting difficult/large/forced transitions. At La Forge @profbof experimented with many things that resembled this feature (and it always surprised me how well they worked). This is the first time I've actually "done the maths" on it though!

How can we measure the cost of transitioning between two different animations? In this blog post I propose a solution based on the assumption of inertialization blending - involving (as usual) some playing around with spring equations! theorangeduck.com/page/inertiali…

A couple years ago I was dancing on 'uptown funk' in a mocap suit with @anorangeduck and @_Quazi_. The goal was to find a way to play any move on a non-cheating ragdoll (or a real robot). Good job guys. It worked.

I've just added a basic implementation of Learned Motion Matching to my previous Motion Matching demo. Hopefully useful to anyone working on this stuff. Results aren't visually any different (as intended), but still cool to see LMM running in the browser! github.com/orangeduck/Mot…

Supertrack: train a ragdoll to balance and play hours of difficult mocap by gradient descent on smooth/learned/differentiable physics. This changes everything.

.@ZivaDynamics presented a machine learning-based system that converts character head meshes into real-time puppets capable of expressing over 72,000 facial shapes within an hour. The video is a real-time demo running in UE5 80.lv/articles/ziva-… #animation #3dart #gamedev

I really can't tell if this is real or fake. That's a milestone.

Another article - this time on joint limits, and how we can automatically compute them from data. theorangeduck.com/page/joint-lim… I've also added interactive web demos for this article and my previous article on code vs data driven displacement - much more fun to play with!

@olafurw You should also mention how redondancy can help a lot.

Which is why TCP behaves this way. Makes sense when explained with that analogy. But now lets talk about UDP packet ordering. 🧵 6/17

@olafurw You should also mention how redondancy can help.

@anorangeduck This code will be very useful. And it's MIT licensed, so we're allowed to copy paste it anywhere. Thank you Daniel you are the MVP.

The demo from the blog post also includes a very basic motion matching implementation somewhat in the flavor of some of our previous research so it may be interesting for people who want to try implementing this technique themselves. github.com/orangeduck/Mot…

Another blog post on animation programming fundamentals, and today it's a big one - code vs data driven displacement - or how we can move a character around a game world in a way which is both responsive and realistic. theorangeduck.com/page/code-vs-d…

This blog post is an important contribution. Go read it.

@anorangeduck Turns out I just live coded a simple motion matching setup. Trying to keep it extensible but still cache friendly. youtube.com/watch?v=jcpIrw…

@anorangeduck Motion matching is a nice example where cache is the most important thing to keep in mind. It is not the place to get fancy with dynamic arrays. Candidate frames are just rows in a big rectangle of numbers, and gameplay searches for the closest row.

IMO the two most important transformations for performant code are Array of Structs to Struct of Arrays and Loop Fusion to Loop Fission. The 1st is essential for cache perf and the 2nd breaks data dependencies and often unlocks more optimizations that were not obvious before.