Robinerd

This might be the most advanced thing I have ever built. So many parts all coming together

@_theshash @broodsugar @Auki If the cloud is very near to you, Maybe it is a mist?

If you didn't catch Robotics Livestream Ep. 2, here's what the panel actually disagreed on - and where they landed. The humanoid question split the room. @broodsugar from @Auki was the most bearish: "I am yet to find a real customer that actually wants to buy or rent a robot with legs." He compared Figure to Magic Leap - $500M raised on a vision that took a decade to even partially materialize. @dabblerer_ from @BitRobotNetwork pushed back gently: humanoids are the most attention-grabbing form factor and the most relatable for people. Specialized robots might ship first, but generalized robots could also find product-market fit if the development catches up. George from @tashiprotocol sided with Nils: "Deploy 10 single-use robots that all talk to each other instead of one humanoid who gets his ass kicked the moment it leaves the lab." On cloud vs. edge, the panel converged fast. Nils was absolute: cloud-controlled robots will never work in production. Customer sites have bad internet and millisecond latency matters for safety. @0xPravar from @nunet_global agreed but added nuance, it won't be pure edge either. The answer is hybrid architecture with a compute layer that doesn't exist yet. Nils coined a term for it: "domain-side compute." Hyper-local compute on the same network as the robot, plugged into the wall. He believes connecting to this will become as common as connecting to wifi. On data, the panel agreed the industry is still early. Nvidia's Sonic model used 700 hours of motion capture. Nils called it less than $100K worth of data. Karen noted it was mostly collected in Nvidia's controlled offices, the real challenge is real-world environments. But both agreed open-sourcing is the right move because it gives teams a foundation to build on rather than reinventing the wheel. On the biggest opportunity nobody's talking about: Nils laid out a thesis that the internet itself needs to grow three new dimensions for physical AI - sensors, spaces, and actuators. George connected it to traffic: autonomous vehicles coordinating at low latency in Beijing would recover the equivalent of pyramid-building time every week. The consensus: the infrastructure layer for physical AI is where the real value is. Hardware will take care of itself. The coordination, compute, and data layers are the bottleneck. Catch the full Robotics Livestream Ep. 2 on YouTube: youtu.be/jUBUnXiaHjU

@Cyn_Cyb3r071Qu3 @CyberRobooo Exactly!! Or even a cute pixar robot. Way too uncanny. Less human is the way

@CyberRobooo I cannot understand why people would rather see this than a robot that looks like a robot.🙄

A humanoid robot just started a job at Ericsson. Not in a factory. In the lobby. Realbotix delivered a new AI vision system called Vinci along with a humanoid robot to Ericsson’s Imagine Studio. It can recognize faces, read expressions, and understand what’s happening around it in real time. So it doesn’t just respond anymore. It pays attention to people. Its job is simple. Greeting visitors. Guiding people around. Helping explain a few things to employees and guests. The first real roles for humanoid robots might not be just physical work. They might be human interaction.

@audrlo If everyone did the extra mile is not "extra" anymore though :P

people who go the extra mile will always win in the end. the problem? most people don't.

@ChrisXswe Very good! Perhaps rumours of improving x algo is true? Hopefully this little dialogue here lets the algo know our interactions are to be prioritised

@RobinerdGames Whaaaaaa..t! Well, it’s food to be back in your feed ✨🤭 How you’ve been Robin?!

Waking up early, I am starting to get a bit tired of social media, lot’s of bots, I feel we’re overall get less and less genuine conversations. I do love the ones I have and I really enjoy the ones I follow who share their hobbies, streamers and composers within VGM.

@MakerInParadise @Tronicssystem @shazcodes If they didnt tell it to NOT remove the codebase doesnt mean they told it to do so. It highlights that you still really cannot trust the AI without oversight, still quite far from "did what it was told to"

Only that’s not at all what happened… an employee allowed their Kiro system to run without oversight and it made a fix by wiping and rebuilding a code base in current use. It did exactly as it was told. Thats not a hallucination, it’s insufficient context delivery mixed with no oversight.

Our ceo fired the entire 12 person QA team last month and replaced them with an ai automated testing pipeline to save $1.2M today, we lost $6M in orders because a bot hallucinated a discount code that made everything in the store 0. The best part? he asked the lead dev to hop on a call with the fired QA lead to see if he’d consult for free to fix it. corporate greed is a mental illness

@BukeBeyond @janusch_patas Depends on use case. For example, performance, portability, compatibility with existing content, mixing with other simpler 3D gizmos or non photo content

@RobinerdGames @janusch_patas but why use triangle solids at all? gaussians have multidirectional light emission (spherical harmonics). we can relight them now, and even render them as solids for shadow casts and refraction. solid extraction can be used for special cases, but often the results are noisy.

From Blobs to Spokes: High-Fidelity Surface Reconstruction via Oriented Gaussians TL;DR: Gaussian Wrapping interprets 3D Gaussians as stochastic, oriented surface elements and derives closed-form vacancy and normal fields, enabling fast, watertight, and compact mesh extraction of full 3D scenes. Contributions: – We introduce Oriented-Gaussians and their associated training strategy in the multi-view setting. – We derive a theoretical connection between 3DGS and implicit surface reconstruction by formulating Gaussians as oriented surface elements, inspired by Objects as Volumes [39]. Importantly, this leads to closed-form expressions for both normal and occupancy fields at arbitrary locations without any additional learnable parameters. – We propose Primal Adaptive Meshing, a mesh extraction procedure that leverages the derived Gaussian fields to produce high-quality, water-tight meshes at controllable resolution, enabling recovery of extremely thin structures such as bicycle spokes (Fig. 1).

Sure but can still use gaussians for where it excels, like anything that doesn't map well to solid triangles. Using gaussian splats alongside traditional meshing or custom shaders and vfx is the ultimate combo, no need to force on technique on everything anyway. Imo this is really big for many reconstruction use cases where you want something realistic from video capture without necessarily forcing a 3DGS render pipeline

@janusch_patas bicycle spokes are not "extremely thin". try hair, which gaussians excel at.

@rowancheung Unsupervised learning via self play was huge many years ago already, deep mind beating the pros at video games and even before that alpha go. The main thing new here is that the simulator actually has to transfer to real hardware correctly which is not trivial.

Researchers just taught a robot to play tennis. From just clips of a few amateur players performing basic forehands, backhands, and shuffles... ...a robot learned one of the fastest, most coordinated physical skills there is. Insane!

@broodsugar Luckily... so far! 🦾😇

I don't think the average humanoid realizes how easily they can break a human.

@Auki @Intercognitive Humans are just subconsciously really good at reasoning about spatial surroundings and also projecting from what others see from their point of view, we don't even realize it is a shared coordinate frame.

Robot coordination is only possible when they agree on a coordinate system. As humans, a shared coordinate system is something we take for granted, but robots have no such thing. Auki and @Intercognitive are solving that. If you missed the X space, catch up below!

@lichtfeldstudio What! Nice, I need to try this now

New densification strat just dropped and will be in the next nighly. Working title: LFS 🙂 A hybrid of IGS+ and MCMC, combining structured updates with mcmc-like exploration. Metrics on mipnerf360 are similar… but as we all know, metrics on toy data are only half the story.

@clankrmedia @broodsugar Both. Real question is how we make those align?

@broodsugar Exactly, but here's the real question: will robots build those cities to serve us, or to serve themselves?

You can just build cities that build robots that build cities.

@Scobleizer @Auki And if mobility in uneven terrain is a need then just wheels on the legs. Still superior. Nature just didn't manage to hit the right order of random selection to evolve wheels...

@Scobleizer @Auki Just like being able to stand idle without a constant flow of micro compensations for balance. I have no numbers but feels significant enough. Last time I parked my car it stayed there with just power off, nice thing about wheels!

Wheeled humanoids >> humanoids in most situations where you want actual work done. Our first one is on its way.

@Scobleizer @Auki Yes, which is a huge deal not much talked about in media so far (compared to degrees of freedom and what not)

@dyor21 @Auki And maybe with a dreamy glow around the edges, glorious!

@Auki By the looks of it, the robots implemented with @Auki get to see the world better than i do without my glasses

Look how clean our renders are getting! So satisfying. 😄

Starting to get pretty good splats now, from custom iOS captures! @lichtfeldstudio

@DanyBittel @NoThanksHoney1 @janusch_patas Have it on our Polestar 4 as well. Similar stitching artefacts but it is very useful to help understand the orientation of car while parking or turning around narrow places. In addition to mirrors and rear cams etc. That car is pretty wide too so it is helpful, (and cool :P)

@NoThanksHoney1 @janusch_patas It is! With lot's of artefacts.. even the left view, the cars jump in and out of view. Not something I'd expect on consumer software.. I guess it's still useful.

No doubts: German car engineering is a dead man walking!

@janusch_patas @Auki Yeah true, this 360 one was trained by a colleague, didn't check the training params earlier. And yeah definitely interesting to try out the densificstion plugin, and others!

@RobinerdGames @Auki Ok, but that one is trained with ppisp as I see. You could also try the densification plugin. It is in the market place which is included in the more recent builds (nightly builds for Windows).

Creating digital twins with phones is great because everyone has a phone, but creating digital twins with 360° video drones is faster and more fun. Here's a gaussian render of our demo store made from Insta360's Antigravity A1.