Physical AI

3D for everyone, everywhere: today, we’re proud to introduce OnTheFly.

https://onthefly3d.com

Born from the GraphDeco research group at Inria, OnTheFly is building the next generation of 3D media tools by putting high-quality 3D capture directly in the palm of your hand. Our ambition is to make capturing, navigating, and editing 3D content accessible to anyone, with no specialized hardware, no expert workflow, and no post-processing delay.

Photography was not always effortless. It once required technical expertise, careful preparation, and long processing times. Today, anyone can take a photo on a smartphone and see the result instantly. 3D is still in its film era.
Professional 3D capture often still requires dedicated hardware, skilled operators, long processing times, and a leap of faith: you only know whether the capture worked once the reconstruction is complete.

OnTheFly makes high-quality 3D capture intuitive, immediate, and accessible on any smartphone.

As you move through a space, the 3D reconstruction appears directly on your screen. You can see what has been captured, identify what is missing, and know immediately that you have the result you need. No blind capture. No post-processing stage. What you capture is what you see on the screen.

Founded by Anthony Schoofs, Andréas Meuleman, Camille Montemagni, and George Drettakis, OnTheFly builds on years of research in 3D reconstruction and novel-view synthesis. Our goal is to bring advanced 3D technology out of the lab and into real-world workflows across construction, manufacturing, marketing, and entertainment.

Our latest research will be presented at #SIGGRAPH2026, marking an important step toward immediate, robust, and scalable 3D reconstruction. We’ll also be exhibiting at the Cap Digital Pavillon France.

Come meet us and see what the future of accessible 3D looks like.

Source: https://www.linkedin.com/posts/siggraph2026-ugcPost-7483319158542155776-smnq

Meet the first Unreal Engine plugin built for billion-scale Gaussian Splats — LCC plugin for UE, now available from XGRIDS.

Import, edit, relight, and render large-scale 3DGS scenes directly in UE, and bring real locations straight into your virtual production pipeline.

Getting your assets in is just as easy: compatible with LCC2, LCC，PLY, SOG, and SPZ Gaussian formats, so your existing 3D assets import into UE and you're ready to create. Whether you work in film and virtual production, digital twins, simulation, games, or XR, it slots right into your workflow.

Five core capabilities:
• Relighting, now closer to reality than ever
• Cinematic depth of field
• Seamless 3DGS + mesh integration
• Built for virtual production workflows
• Stable high frame rates, smooth VR experience.

Capture the real world. Build the world model.
Download👉: https://bit.ly/4yF4nVZ

Join our Discord Community here!
https://lnkd.in/g9JAeexK

🎥 Video Credit to Active Retech & Lukasz Mirocha, PhD 🔜 SIGGRAPH LA

#UnrealEngine #3DGS #GaussianSplatting #VirtualProduction #RealityCapture #SpatialIntelligence #XGRIDS #UE5

Source: https://www.linkedin.com/posts/unrealengine-3dgs-gaussiansplatting-ugcPost-7483068297115144192-481A

Exciting news: Reinforcement Learning (RL) coming soon to the Telekinesis Agentic OS!

Here's a first look at a robust locomotion and stair-climbing policy trained with PPO.

The Telekinesis Agentic OS aims to bring together the key technologies of Physical AI into a single ecosystem.

Our next focus is to release the RL training stack with:
1. Key on-policy algorithms such as PPO
2. Distributed training across multiple parallel environment
3. Native support for NVIDIA Omniverse and MuJoCo
4. A growing suite of humanoid training environments

Our long-term goal is simple: give robotics developers a single OS that brings together the core building blocks of Physical AI, from vision and motion control to reinforcement learning, VLA models, World Models, and beyond.

Explore the Agentic OS docs: https://docs.telekinesis.ai/
Join our Discord of robotics and computer vision engineers: https://discord.gg/syc5TqfsSK

#robotics #physicalai #ai

Source: https://www.linkedin.com/posts/suman7495_robotics-physicalai-ai-ugcPost-7482696382894542850-szPB

A useful way to think about robot training data.

Many teams building VLA models talk about data volume, but fewer explicitly decompose robot data into layers to discuss what each layer is actually useful for.

I spent three years scaling teleoperated data collection at Yango Robotics, from about 15,000 episodes a month to 400,000. Along the way, I learned that raw episodes / hours count can be often a misleading metric.
From my experience, data and labeling quality matter far more than people expect. A massive dataset only helps if it carries enough diversity, the right signal, and precise labels for the training stage you're targeting.
The real question devs should be asking is where your data sits on a crucial tradeoff: scale vs. contact fidelity.

Check out the video below for examples of the different data types, including the ones we make at Toloka.

Here is what that data spectrum looks like in practice:
Egocentric Video - great for scale, diversity, task context, and high-level planning. The limitation? It lacks ground-truth hand pose and a strong contact signal. You can estimate hand pose after the fact, but the quality rarely holds up for contact-rich manipulation on its own.

UMI-Style Data - costs more to collect, but gives you real 6DoF end-effector pose and a camera angle that actually sees contact. It sits right in the middle of the spectrum, some variants lean closer to egocentric collection, others closer to real teleop.

Real Robot Teleop - the most expensive and least scalable. It is also the only data that removes the sim-to-real and human-to-robot embodiment gaps entirely.

TLDR: If you want to teach a specific robot a specific task, the most direct path is usually to collect teleop data as close to that exact task, robot, environment, and action space as possible.
But if your goal is a generalist robot model, you probably need a layered approach: scalable egocentric data for breadth, higher-fidelity gripper data for contact, and real robot teleop for grounding.

Curious if you are thinking about this differently. Let's discuss in the comments.

https://www.linkedin.com/posts/v-toropov_a-useful-way-to-think-about-robot-training-ugcPost-7482415453257134080-rMB3/?utm_source=social_share_send&utm_medium=member_desktop_web&rcm=ACoAAAYuSGUB01rbMyTFNW4SkTf50dynCH7Luuw

https://github.com/kabilankb/so101-bench-groot-n1.7

Fine-tune and evaluate GR00T-N1.7 on SO-101 Bench in an Isaac Lab digital twin — no robot required.

This is the GR00T-N1.7 variant of 5hadytru/so101_bench: a simulated twin of a language-conditioned SO-101 tabletop manipulation benchmark (bin / next-to / between / move tasks over 56 household objects), retargeted from GR00T-N1.6 to N1.7 (Cosmos-Reason2 / Qwen3-VL backbone). See the upstream repo for the benchmark design and the paper's real-robot findings.

Can Cosmos 3 Simulate Robot Behavior? One Starbucks Scene, a Dozen Robot Behaviors

We set out to answer one question: can you use NVIDIA Cosmos 3 to simulate a whole range of robot behaviors from a single image — not one demo, but a repertoire? So we treated it as an exploration, built from one Starbucks scene, and chased every interesting thread it opened. Here is the full story: the use cases we tried, the prompt strategy that changed everything, the model annotating its own output, the loop we closed, and what it all says about where world models actually fit in physical AI.

https://www.linkedin.com/pulse/can-cosmos-3-simulate-robot-behavior-one-starbucks-scene-mohammadi-k2rkc/

https://www.nvidia.com/en-us/glossary/simready/


SimReady (simulation-ready) is an OpenUSD-based framework that defines the physical properties, semantic labels, material attributes, and 3D asset metadata needed for physical AI simulation workflows.