Raphael Levy

Head of Europe @Foxglove. Physical AI & Robotics. Paris based, SF roots 🇫🇷🇺🇸

[Raphael Levy @Raphael_Levy_]

@mbyBelyy Yep!

[Raphael Levy @Raphael_Levy_]

Showing up at #GTCBerlin!

NEURA Robotics just opened a 2,300 m² gym in Munich where companies teach robots their jobs. And unveiled a new humanoid to do the work.

The TUM RoboGym, built with the Technical University of Munich, has been training real workflows since September. Handling cardboard boxes, filling bottles, loading wooden components. Partners from more than ten industries, from automotive to pharma to agriculture.

The humanoid is 4NE1 Gen 4. 20% lighter, slimmer, with a black textile skin and better hands. The detail I like most is the Smart Limbs: reflexes built directly into each limb, so it reacts on its own instead of waiting on the brain. It’s designed for series production and runs on Neuron AI, which takes touch, force and sound into account alongside vision.

The idea behind all of it is simple. German companies already hold decades of process knowledge. Bring a task to the gym, the robot learns it, and the skill spreads to every connected robot.

I posted about the Aachen gym a few months ago. Munich is the next one, with Europe, the US and China to follow.

It’s a very German way to build physical AI. Start from the factory floor, not the demo.

[Raphael Levy @Raphael_Levy_]

@microagi 1PB per day? Waouh!

[Raphael Levy @Raphael_Levy_]

@lukas_m_ziegler The goat!

And Jensen

[Raphael Levy @Raphael_Levy_]

Nebius open-sourced a Physical AI Workbench this week, built on NVIDIA's stack. And @Foxglove is in there, for observability.

That line made me grin! Doesn’t matter whose cloud or whose GPUs sit underneath, when engineers need to see what their robot actually did, they reach for us.

It’s also a good signal for Europe. Look at who’s already building on it: Moon Surgical, Bellboy Robotics, Sereact. These aren’t hype names, they’re teams shipping fleets and surgical robots into the real world, solving unglamorous problems nobody talks about.

I learned this in martial arts before I learned it in robotics: you don’t improve by doing more reps, you improve by reviewing the tape and fixing the one thing you got wrong. That’s the layer @Foxglove owns. Good to see the industry finally building around that instead of treating it as an afterthought.

https://t.co/fg3X69EaJM

[Raphael Levy @Raphael_Levy_]

@DutchMartin @superintel_xyz @pollenrobotics @NVIDIARobotics Hahaha

[Raphael Levy @Raphael_Levy_]

Switzerland’s largest pre-seed ever just went to robots that can feel. $30M, Zurich, announced today.

Enact was founded in March. Founder Manuel Yves Galliker led the team at 1X that taught NEO to walk, then worked at Physical Intelligence. Now he’s building a nervous system for robots: force sensing across the whole body, down to the fingertips, with a model that turns those signals into movement.

Most robots still rely almost entirely on cameras. Plugging in a cable, fitting a seal, starting a screw that catches is still manual work. In Germany only 4.6% of small and mid-size companies use robots, versus 24% of large ones. Enact is going after that gap: the last mile of assembly, many variants, small batches. “China wins on volume; we win on variety.”

12 people from Google DeepMind, 1X, Amazon Robotics, Helsing and Physical Intelligence. Yann LeCun and Arthur Mensch are in as angels, alongside redalpine, Notion Capital and firstminute capital.

Touch is having a moment, and Zurich is the robotics capital of Europe. I’ll keep dying on this hill.

[Raphael Levy @Raphael_Levy_]

Jensen Huang just knelt down on stage in Berlin to say hello to a 25 cm robot.

That’s Microduck, Hugging Face’s open source biped. His line to it: “Did you know that you were created in here?” He meant the simulation. Microduck learned to move inside NVIDIA’s sim stack (Newton physics, Omniverse) before it ever touched a floor.

Now look at who’s standing behind him. Disney Research robots. Enchanted Tools. A whole row of robots on the #GTCBerlin stage.

His message to Europe: “There’s no question everybody will become an AI company and we’re going to have a thriving AI economy here in Europe.” He also said Europe’s manufacturing and mechatronics base could lead in humanoid robotics.

I’ve spent this year in rooms with the people building these robots in Paris, Zurich and Munich.

I’m not worried about Europe. I’m impatient for it.

[Raphael Levy @Raphael_Levy_]

@bercankilic Billion robots in the making!

[Animesh Garg @animesh_garg]

Two new egocentric benchmarks from @microagi evaluate VLM readiness for robotics:
MicroPlanBench and MicroAnnotationBench.

tl;dr: Semantic reasoning ≠ perceptual timing
A lot of work to do.

MicroPlanBench (consequential next-action planning)
When predicting the single valid next atomic action from 8s of egocentric video, SOTA models still fail ~1/3 of the time (GPT-6 Astra: 65.6%, GPT-6.1 Sol: 65.5% at 1/5th cost).
- GPT-6 Astra error breakdown: 60% wrong manipulation, 23% wrong hand assignment, 17% wrong object/destination.
- Across all flags, 42% of failures involve tracking the wrong hand.
https://t.co/Yd6cvJSiuh

MicroAnnotationBench (continuous temporal grounding)
Knowing what happens ≠ knowing when
Continuous 60s+ dense manipulation video remains unsolved for generalist VLMs:
- GPT-6 Astra/Sol over-segment (~22 segments/clip vs. 13.8 human mean), collapsing precision (~39%) and strict Boundary F1 (<20%).
- Scaling Test-time compute fails to fix perceptual boundaries: scaling Gemini 4 Argon to high reasoning effort dropped its headline score by 4.9 points while costing 4.7x more
https://t.co/LbEfZblJbJ

Currently, Small specialists match/beat frontier APIs at scale.
@microagi's self-hosted nanoannot-1 outperforms the frontier models (Gemini 4 Argon) across the board:
- 54.8 vs 49.6 overall;
- 32.9 vs 27.3 on strict ± 0.667s Boundary F1
- ~1/10th the cost (~$0.30 vs $3.14/video hr & >300x cheaper than GPT 6 Astra)

Solid work by @ErguzelYag, Ömer Adanir & Team.

[Raphael Levy @Raphael_Levy_]

@adityabhatia89 Precisely! Robots run on data :)

[Raphael Levy @Raphael_Levy_]

A robot cleaning a toilet. Then washing and drying its own hands!

That second part is what is different. It’s not just completing the task. It’s completing the hygiene protocol that comes after the task. That’s a different level of awareness entirely.

This kind of deployment is exactly what Reflex Robotics is building for. A wheeled humanoid designed to handle end-to-end jobs in real operations, zero integration, works with your existing carts and equipment, and gets more capable every shift. Built by a team from Anduril Industries, Boston Dynamics, Waymo, Skydio.
Backed by Khosla Ventures.

The unglamorous use cases for physical AI are moving faster than the glamorous ones. Hard to staff, high turnover, physically demanding, hygiene-critical.

The business case is crystal clear.

[Raphael Levy @Raphael_Levy_]

Unitree is back with a new banger.

Meet the H2-X “Kung Fu Legend Guardian”, announced today.

In the demo video it kicks a car door until it deforms, lifts a car with its bare hands, and backflips off a 4 meter platform before landing clean. Unitree’s own advice for the humans nearby: stay 2 to 3 meters away.

The detail I care about more than the kung fu is the battery. About 8 hours, up from roughly 3 on the base H2. That’s a full shift, which is what turns a stunt robot into a patrol, security or emergency response robot.

Specs and pricing aren’t out yet, so I’m curious what the runtime looks like with real perception and patrol workloads running. But Unitree keeps shipping at a pace nobody else matches.

We got kung fu humanoids before dishwashing ones. I’m not complaining.

[Raphael Levy @Raphael_Levy_]

Eka Robotics just raised $150M at a $1.45B valuation! Andreessen Horowitz led. Announced today.

The number that stopped me: 1,500 operations per hour. That’s 2.5x the human baseline of around 600. Not matching humans. Beating them.

Their thesis is the most interesting I’ve read in robotics this year. Everyone else is training on human demonstrations. Eka is doing the opposite. They believe mimicking humans caps your ceiling at human speed. To go superhuman, you need to master physics, not imitate humans.

They built a Vision-Force-Action model trained entirely through reinforcement learning in simulation. The precedent they cite: AlphaZero, which beat every human chess player without a single human game in its training data.

The framing from their blog: “Generality without performance isn’t a product. It’s a lab demo.”

Hard to argue with that.

Wired called their approach “a ChatGPT moment for the physical world.” a16z’s GP called it “the first approach that truly solves manipulation.”

Something is happening in physical AI right now. The pace is relentless.

[Raphael Levy @Raphael_Levy_]

I used to move furniture for a living as a student.

35kg. That’s a big box. A small washing machine. The kind of thing that ends your day early when you drop it wrong.

Korea Advanced Institute of Science and Technology’s Humanoid v0.9 just pushed a 35kg load and lifted 20kg. Coordinating its whole body while doing it. Legs adjusting for balance, arms applying force, posture managed throughout. Not two separate tasks. One fluid movement.

The demo shows it arranging objects to spell “KAIST” using onboard visual feedback. Trained with reinforcement learning and human demonstrations.

I remember learning very quickly that you don’t lift with your arms. You use everything. Legs, back, core. The whole body.

Turns out that’s also how you build a humanoid that can actually move things.

[Raphael Levy @Raphael_Levy_]

I am excited to share that @foxglove is teaming up with Women in Robotics in Paris at Enchanted Tools on November 19th.

It’s an evening bringing together women building in robotics, engineering, and entrepreneurship, with talks from three women leading robotics and software teams: Celine Etcheverry (R&D Software Development Director at BALYO), Lise Delobel (Project Coordinator at Enchanted Tools), and Marine Petriaux (Head of Manipulation at Wandercraft).

Great chance to connect with some of the sharpest people working on robotics in Europe right now!

Registration (with host approval) here: https://t.co/6b10uopdia

[Raphael Levy @Raphael_Levy_]

@adityabhatia89 It's a good point, and that's why autonomy is an important part of the stack + data is super key here

[Raphael Levy @Raphael_Levy_]

Dream Labs just came out of stealth with $90M!

Dream-1 is a robot control model that works on robots it has never seen, doing tasks it was never trained on. Zero-shot. You plug it in, you run it. $3 per hour.

They tested it on 27 different robot bodies in simulation. The PR2, an 18-year-old research robot. New Chinese arms. Hugging Face’s $399 Microduck. It controlled all of them without any embodiment-specific data.

The technical insight: instead of outputting joint angles (which lock a model to one robot), Dream-1 predicts where each gripper should be in space. Every bimanual robot shares that interface. Its own IK solver handles the translation.

The bigger insight is what this does to the economics. Robot hardware is getting cheap fast. 7 semi-humanoid robots under $15k. A capable bimanual setup for $2k. If a model can make any robot useful from day one, the flywheel can finally start.

Dream-1 also outperforms Opus 5.5 and Gemini 4 Argon on RoboDojo benchmarks. At 16x lower cost than Gemini 4 Argon.

This is a big one.

[Raphael Levy @Raphael_Levy_]

NEURA Robotics just moved into cleaning.

They acquired ADLATUS Robotics GmbH and launched a new product family: autonomous floor cleaning and sweeping robots for production floors, warehouses and commercial spaces.

NEURA started with humanoids. Now they’re building a full automation portfolio. Material transport, autonomous cleaning, and more to come.

The strategy makes sense. Once you’re already deployed in a facility with your robots, the question becomes: what else can you automate in that same space? Cleaning is everywhere, hard to staff, and runs on predictable schedules.

This is what a robotics platform company looks like as it scales. Not one robot. A full suite.

[Raphael Levy @Raphael_Levy_]

@olegzanx Yes, that's the idea! You build a product with an idea on how to sell it. That's what makes it scalable!

[Raphael Levy @Raphael_Levy_]

~30 hours, real hardware, a room full of Europe's sharpest robotics builders!

That was the ERH Physical AI Hackathon in Zurich this weekend. Robot arms, mobile platforms, sensors and compute, all handed to teams on arrival, with mentors on site and a jury of founders, researchers and investors waiting at the end.

Spent part of the weekend mentoring teams on GTM at the European Robotics House, helping young robotics builders think through who their customer actually is before they fall in love with the tech.

Huge respect to ESRA and the host team for making it happen. This is what the European Physical AI ecosystem looks like when it shows up in person!