Showing off some more dexterous skills with grippers. 🦾
Watch Gemini Robotics 2 x Franka Robotics 's FR3 Duo in action.

Learn more about the release here: https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/

Source: https://www.linkedin.com/posts/nicolas-heess-5600922b8_showing-off-some-more-dexterous-skills-with-ugcPost-7489709302006390784-m8Dt

Finally finished a complete implementation of Gaussian Splats that behave like a 3D mesh in Unity. They can cast and receive shadows. Receive or cast lighting. Integrate completely with any mesh in scene. They can even be raytraced. Cast reflections. Use Ambient occlusion. They even support complex shader effects by using #Amplify shaders.
They will be integrated into my new Deckard Render Studio. They can be enhanced with Deckard Render for photorealistic look offline renderings. Or used with depth aware image effects like DOF when using DLSS.
The approach is similar to Stockasting gaussian splats approach. Tailored for Unity.
Gaussian Splatting

Source: https://www.linkedin.com/posts/oliver-pavicevic-ba67a651_amplify-ugcPost-7489783617972305920-bx8T

At Mirador 3D, we are building a Unity plugin for large Gaussian Splatting scenes using PlayCanvas .SSOG format, designed to stream millions of splats across multiple platforms .

We have also integrated Meta XR hand tracking and gesture-based locomotion on standalone Meta Quest 3, while working within a limited active-splat budget.

Our next steps are to validate performance on Meta Quest 3, further optimize streaming, memory usage and rendering, and expand support for additional platforms.

#Mirador3D #GaussianSplatting #3DGS #MetaQuest #Unity #XR #PlayCanvas #SuperSplat

Links:
https://superspl.at/scene/8429e5e2

Source: https://www.linkedin.com/posts/jerson-march%C3%A1n-jimenez_mirador3d-gaussiansplatting-3dgs-ugcPost-7488896045683392512-jRcW

We ran an experiment at a live show at the California Theatre: find the 3D location of every photo taken during the event.

Live shows are one of the hardest environments for a VPS. Low light. Smoke. Strobes. Almost every photo pointed at the performer. Hundreds of people moving through the frame.

MultiSet AI VPS Gen2 localized every photo in real time.

Some of these images have almost no visible features. A human looking at them can't tell what's going on or where they were shot. Gen2 can. That's the direction we're pushing with Gen2: capability past human vision, closer to superhuman.

What it means in practice is that every photo and video captured in the field is spatially ground truthed. You can go back and revisit them in place. An early version of spatial memory.

Try it today: developer.multiset.ai

Source: https://www.linkedin.com/posts/niksawlani_we-ran-an-experiment-at-a-live-show-at-the-ugcPost-7493258612077895680-Cmnf

"Teleop data is a couple classes above egocentric data.”

It was great to hear Lukas M. Ziegler mention Adamo on etn. a couple days ago, and I completely agree with his point.

This is fast becoming an imperative conversation in the robotics space, as hardware companies are starting to realise where teleop data sits on the chain of data collection, in order to deploy their robots.

Teleop data is at the top of the pyramid.

I agree with Lukas that egocentric data is just simply not enough, and uses the example of chopping onions as an example of this.

When a human makes a mistake you will have to post filter the potential mishaps or mistakes.

At Adamo we’re continuing to scale our teleop data collection services, collecting unique and high quality data for your robot to deploy seamlessly into the real world.

Source: https://www.linkedin.com/posts/james-naylor98_teleop-data-is-a-couple-classes-above-egocentric-ugcPost-7493608178887282688-3KdW

https://docs.nvidia.com/learning/physical-ai/getting-started-with-isaac-for-healthcare/latest/agentic-workflows/index.html?source=everyonesocial&network=linkedin&userID=09b33672-8dd2-4ecb-b8a5-862096d6faf9&es_id=cee90c7bef

Agentic Workflows in Isaac for Healthcare
From prompt to policy
Prompt Your Way to a Healthcare Robotics Simulation
Isaac for Healthcare supplies the simulation, data, and policy building blocks for robotic autonomy. In this course, you use agentic workflows to create a healthcare task, collect demonstrations, train a policy, and validate it in Isaac Sim.

Source: https://www.linkedin.com/posts/thomas-richter-trummer-72939014_gaussiansplat-3dgs-gaussiansplatting-ugcPost-7495054399787126784-qI_c

A single static #GaussianSplat capture can now become an animated, relightable character.

This is different from 4D Gaussian Splatting, where a captured sequence is reconstructed and replayed.

Here, there is only one static scan (by 3D Scan Studio iris).

The character is rigged once, arbitrary Mixamo-compatible animations can be applied, and the same rigging data is used for relighting.

The relighting happens per splat — without a proxy mesh.

Normals from the rigging mesh are transferred onto the splats and used directly during rendering. This helps preserve fine details, especially around faces, where a separate proxy surface can easily diverge from the actual splat representation.

And there is no additional relighting setup.

Once a humanoid splat is rigged, the data needed for relighting is already there.

capture once → rig → animate → relight → reuse

The same scan can then be used as an avatar, an NPC, or with completely different animations and lighting conditions.

We built this into the existing humanoid splat pipeline in Arrival.Space

Scan data by 3D Scan Studio iris / Steam Studio, Japan, and UVRSE.AI
shared on SuperSplat. Their exceptionally clean capture made it a great dataset for testing this.

#3DGS #GaussianSplatting #Web3D #RealTimeRendering #3DScanning #PlayCanvas

Source: https://www.linkedin.com/posts/thomas-richter-trummer-72939014_gaussiansplat-3dgs-gaussiansplatting-ugcPost-7495054399787126784-qI_c

Robotics deep dive: WHAT SIMULATION CAN'T TEACH

Simulation solved walking.
A humanoid learns to walk in about fifteen minutes on one gaming GPU, across a thousand randomized copies of itself, and the policy transfers to a real robot with no retraining. Layered randomization cut fall rates on Agility's Digit by 40%.
That is solved in the way engineering means solved: cheap, repeatable, boring.

Now point the same machinery at a hand.

Stanford's TRANSIC ran the standard domain-randomized recipe on real furniture assembly, transferred zero-shot.
Averaged over four tasks it succeeded 36% of the time.
On insertion, the one task that is nothing but contact, it managed 15%.

Same technique, same era, two completely different outcomes.

The difference is what each problem asks of the simulator.
A walking robot mostly needs the ground to be there.
Contact with the floor is an interruption: large, brief, forgiving, and the policy's job is to recover from it. Get friction roughly right and the gait survives being wrong.

Manipulation inverts that.
Contact is not a disturbance to reject, it is the task.
Seating a peg is a negotiation of forces in the last few millimetres, friction that changes with angle, a part that wedges, a surface that gives.
The TRANSIC authors name the culprit directly: sim-to-real fails on "poor physical realism, where physical interactions such as contact and collision are poorly simulated."

Rigid-body engines approximate the friction cone, let bodies interpenetrate slightly and then push them apart, and integrate stiff contact at a timestep chosen for speed. Every one of those choices is fine when you are catching yourself on a floor and wrong when you are feeling for a hole.

So the field's cheapest win and its hardest wall come out of the same tool.
Simulation gave us legs almost for free.
It has not given anyone hands, because a leg can be wrong about contact and still walk.

Data is a buffet"

Today’s panel had an attention-grabbing title:
“The Data Wars.”

Robotics data has become an increasingly hot topic over the past few months.

More companies are building data collection platforms, exploring new sources of data, and trying to figure out what kind of data will matter most for training the next generation of robot models.

So the “war” for robotics data is very real.

But one panelist Naveen Kuppuswamy offered a different way to think about it.

Instead of focusing on whether one type of data is good or bad, he compared data to a buffet.

At a buffet, there are many things you can eat. They may all have value.

But your appetite is limited.

A robot cleaned our office today.

Thanks Tau Robotics and Cornelia Weinzierl, that was pretty awesome!

Source: https://www.linkedin.com/posts/jamelhammoud_a-robot-cleaned-our-office-today-thanks-ugcPost-7495250619784417280-yK9G

Source: https://www.linkedin.com/posts/olika3d_vray-3dgs-artgineering-ugcPost-7495495294688710656-yE68

Relighting a Gaussian Splat with V-Ray 7.4. 🏡 ☀️

It’s been a few months since my last Gaussian tests, and Chaos just dropped some pretty interesting splat features in their latest #vray update.
Relighting your captures? Definitely had to try that.

First step: I created a new splat with very flat, diffuse lighting so I had a clean base to work with. I also wanted to see how far I could push the details. The 5090 ended up digesting a massive 20-million-splat .ply (about 4.5 GB in #Postshot) without breaking a sweat.
It’s really interesting to see how #3DGS behaves in these setups, where it actually holds together, and exactly when it breaks. The tree shadows, for example, look surprisingly good. But there’s a clear tax on render times once you turn shadowing on.
Need more testing...
Anyone else messing around with these new V-Ray tools yet?
#artgineering

Links:
https://youtu.be/IbefDR3bAeU?si=2Co7xxNzBSpnL1pF

Source: https://www.linkedin.com/posts/olika3d_vray-3dgs-artgineering-ugcPost-7495495294688710656-yE68

Source: https://www.linkedin.com/posts/ronald-van-loon-5411a_physicalai-robotics-enterpriseai-ugcPost-7495372495592767488-qLTO

One of the biggest bottlenecks in Physical AI isn't compute. It’s the quality of human demonstration data.

The race to deploy autonomous robotics usually hits a major bottleneck: collecting clean, real-world training data is slow, expensive, and difficult to scale.

General Intelligence Labs’ new EGO1GS headset takes a fundamentally different approach.
Instead of relying on operators manually teleoperating robot arms for hours, workers wear a 200-gram headset while performing standard physical tasks like sorting, assembly, or handling tools.

It captures stereo video via global-shutter cameras, 400Hz motion data, audio and hand activity, all synchronized on a single hardware clock into one MCAP file. On-device hand detection flags capture issues in real-time, preventing dead datasets from slipping into production pipelines.

The strategic takeaway for leadership:
Physical AI training is expanding beyond tedious teleoperation toward scalable human demonstration.

As this happens, proprietary frontline data, capturing exactly how skilled workers interact with the physical world, becomes a strategic competitive advantage.

The winners in enterprise robotics won't just be the teams with the best models; they will be the ones with the highest-quality real-world interaction data.

Which physical workflows in your organization are poised to become training data for embodied AI?

Video: Interesting Engineering/YT

#PhysicalAI #Robotics #EnterpriseAI #ArtificialIntelligence #AI #Automation #Technology

Source: https://www.linkedin.com/posts/ronald-van-loon-5411a_physicalai-robotics-enterpriseai-ugcPost-7495372495592767488-qLTO

Source: https://www.linkedin.com/posts/the-largest-fully-annotated-open-egocentric-ugcPost-7496556642713468928-dA-q

The largest fully annotated open egocentric human dataset.

Today we're open-sourcing EgoSuite-Open100K with Hugging Face: 100,000 hours of egocentric human data, free for anyone to use.

Physical AI has a scaling law of its own now, and its input is human data. That part is settled. What the field still needs is a shared foundation to build on.

15,000+ tasks. 15,000+ real scenes, from factory floors to retail backrooms. Hand pose, body pose and subtask-level semantics, with wrist camera coverage on part of the set. Licensed for commercial training, not research-only.

We build the data infrastructure for Physical AI. Open100K is the first public layer of it.

The first 10,000 hours are live now, with the rest rolling out in stages.

Project page: https://egosuite100k.lightwheel.ai/

Dataset on Hugging Face: https://huggingface.co/collections/LightwheelAI/egosuite-open100k

If you train on it, tell us where it falls short. That's what decides what we open next.

Source: https://www.linkedin.com/posts/the-largest-fully-annotated-open-egocentric-ugcPost-7496556642713468928-dA-q

Sky Kim
• 2nd
Senior Software Engineer @ Unity | On-Device AI, Agent, XR, Accessibility | Motivational Speaker
16h • Edited •
Connect
Will World Models replace Game Engines in Physical AI?

It was the question I heard most. And it makes sense: Genie 3 generates explorable 3D worlds from a single text prompt, and NVIDIA Cosmos, trained on 200M videos, is closing the sim-to-real gap.

But world models complement game engines, they don't replace them. Non-deterministic control, spatiotemporal hallucination, and massive compute costs remain unsolved. Physical AI demands the opposite: deterministic physics and controllable authoring, exactly what game engines have delivered for decades, from 'Playing for Data' extracting labels from GTA in 2016 to DeepMind training agents in Unity.


Game engines play three roles in Physical AI:

🏋️ Action Space: a safe training ground where AI learns policies through unlimited trial and error in physics-governed worlds

🏭 Synthetic Data: mass production of sensor data, 3D trajectories, and pixel-level labels

✅ Validation: testing high-risk edge cases and multi-agent policies before touching real hardware


Here are four examples. Unity connects with the robotics and ML ecosystem, from ROS and Autoware to MuJoCo and Isaac Lab, making workflows like these fast and easy to build.

1️⃣ Robot Arm Pick & Place. Build the scene in Unity, train with ML-Agents, run inference with Sentis, and deploy to a real robot over ROS-TCP-Connector. A policy trained in the virtual world drives a physical robot arm.

2️⃣ Synthetic Data Generation Engine. Scan hard-to-model spaces with 3D Gaussian Splatting, bring in robots as CAD models via Asset Transformer, and build the scene in HDRP. It then mass-produces RGB/Depth, masks, 3D bounding boxes, and human poses.

3️⃣ Autonomous driving with AWSIM. Mount LiDAR and cameras on a virtual Lexus RX450h in a recreated Nishi-Shinjuku scene, looped with Autoware over ROS 2. The same stack validated in simulation drives the actual Lexus on real roads.

4️⃣ Isaac Lab to Unity. Train RL policies on GPU-parallel MuJoCo Warp physics with domain randomization, export to ONNX, and run them in Unity via the MuJoCo Plugin and Sentis. Training and visualization share the same physics.

And this is already happening in the field. Manufacturing, mobility, healthcare, space, and infrastructure. SEW EURODRIVE cut commissioning time by 30-50%, Medtronic validates surgical robot AI in digital twins, and NXRT trains driving AI in MR simulation. Game engines are already the foundation of Physical AI.


However, Physical AI is still vast and complex, and the barrier to entry feels high. Here is the good news. With the Unity CLI, AI agents can drive the entire engine, and what once demanded a robotics lab now starts with a prompt. You can install the Unity CLI, and build your own Physical AI, simpler and faster than anyone else.

Thanks to the Korea Information Processing Society (KIPS) for having me, and to all the attendees who took the time to join my session on “Game Engines and Physical AI”.

#PhysicalAI #GameEngine #Unity #Robotics