Welcome to Le Physical
Hi everyone,
Welcome to the first edition of our community newsletter. As announced at Physical AI Paris Meetup #3, we’re trying a new way to keep the conversation going: a short digest every two weeks.
Whether you joined us in September, came to an earlier meetup, or couldn’t attend last time, we’re glad you’re part of the community.
Back at La Cristallerie
On 16 September, we met at La Cristallerie in Paris for an evening with Steven Palma, LeRobot Tech Lead at Hugging Face. His talk walked through the robot learning workflow, from teleoperation and collecting datasets to training policies and deploying them on hardware.
Missed it, or want to go back over the details? Steven’s slides are online: Browse Steven’s LeRobot presentation →
Thank you to Steven for sharing his work, to everyone who brought questions and conversations, and to Hexa for the space, food, drinks and coordination. A few moments from the evening:
Physical AI Paris Meetup #3 at La Cristallerie, 16 September 2026.
Steven Palma on control and teleoperation with LeRobot.
Five things worth your time
1. NVIDIA opens a 2,000-asset library for robot simulation
Image source: NVIDIA / Hugging Face
In short: On 21 October, NVIDIA and Hugging Face launched SimReady Central. It’s a free library of more than 2,000 assets ready to use in simulations.
Key points:
The assets include robot arms and grippers, warehouse equipment, racks and household appliances.
Each one comes “sim-ready”, with collision geometry, mass, friction, articulated joints with limits, semantic labels for synthetic data, and materials for sensor simulation.
Franka, FANUC, Universal Robots, Robotiq, Comau and Standard Bots contributed models of their own hardware, so you can use an arm you actually own in Isaac Sim or Isaac Lab.
Our take: Building a simulation scene used to mean modelling your hardware, guessing its mass, friction and joint limits. A library of physically accurate assets provides a shortcut for sim-to-real and reduces its cost.
2. Physical Intelligence releases its models through an API
Image source: Physical Intelligence
In short: On 13 October, Physical Intelligence launched a beta of the Pi API, which lets teams fine-tune and run its latest robot models without managing the infrastructure.
Key points:
You upload your own robot data, run fine-tuning on compute Pi manages, and serve the resulting model through a hosted gateway.
Pi says its latest model fine-tunes with less data and is more capable than earlier ones.
Most of Pi’s early deployments ran inference in the cloud, and some ran into network connectivity problems.
Access is by waitlist and limited to a small group of early companies for now. No pricing has been published.
Our take: If you’ve been fine-tuning the open π0 or π0.5 weights yourself, this is a good managed alternative. The question is whether cloud inference holds up for real-time control.
3. Frontier LLMs take the controls
Image source: Artificial Analysis
In short: Several studies this month gave general-purpose frontier models, GPT, Claude and Gemini models, direct control of robot arms with no robot-specific training. They understand the task well, but their performance in the physical world is still subpar.
Key points:
On 14 October, Artificial Analysis launched AA-Robotics v0.1. Each model drives a real 6-axis arm zero-shot through a simple HTTP API, writing its own code to move it and watching through four cameras across eight tabletop tasks.
GPT-6 Astra leads with 69.4/100, ahead of GPT-6.1 Sol (64.6), Gemini 4 Argon (55.3), Claude Opus 5.5 (48.8) and Claude Fable 5.1 (46.3).
Grasping is the weak spot: even the best model landed only 45% of grasps. Models often declared success too early or knocked over finished work, and a task took about 30 minutes, roughly 100 times slower than a person.
In a separate preprint, a frozen GPT-6 Astra that output every arm waypoint itself reached 92.1% on the RoboMME memory benchmark, well above the best previous leaderboard entry of 60.8%.
RoboColiseum, posted on 20 October, ran four frontier models through a shared control harness. It found near-perfect instruction following but weaker manipulation, with Claude Opus 5.5 best overall.
Our take: For now these models work better as planners and prototyping tools than as controllers. Artificial Analysis itself says they’re far too slow and expensive for production. The question for builders is whether general models will outperform fine-tuned robotics models and where efforts are best spent.
4. Europe’s physical AI push gets a model and a big pre-seed
Image source: NEURA Robotics
In short: In the space of a week, Germany’s NEURA Robotics released a foundation model built around touch and other senses, and Swiss startup Enact Intelligence raised a $30M pre-seed to build robots that feel force.
Key points:
NEURA’s Neuron AI, introduced on 14 October, combines 3D vision, touch, audio, robot state and language with memory. NEURA says it runs on the robot itself across five of its robot types.
NEURA hasn’t published any benchmarks for it yet.
Enact came out of stealth on 22 October to build robots that sense force and touch through their own actuators. Its CEO previously worked at Physical Intelligence.
Enact’s angel investors include Mistral’s Arthur Mensch, Yann LeCun and Hugging Face’s Thomas Wolf.
Our take: Both bets point the same way: touch is becoming a priority for European robot learning. But NEURA hasn’t released any test results for Neuron AI yet, so there’s no way to check how well it actually works.
5. RoboJEPA: Meta scales up robot world models
Image source: RoboJEPA (Meta FAIR / Mila)
In short: On 7 October, Meta FAIR and Mila released RoboJEPA, an 8-billion-parameter “world model” that predicts what will happen when a robot moves. They also published the weights, the training code and the code to run it on a robot.
Key points:
It was trained on about 15,000 hours of video from 23 public robot datasets, covering 12 different robots.
Instead of generating future video frames, it predicts a compact summary of what the camera will see after an action. A robot can then plan by trying out moves in this “imagination” and choosing the ones that bring it closer to a picture of the goal.
The team trained models from 22 million to 8 billion parameters, and their prediction error fell steadily as compute grew. A trend line fitted on the models up to 2B correctly predicted how the 4B and 8B models would perform before they were trained.
On a real Franka arm, with only a goal image and no task-specific training, the 8B model succeeded 67% of the time at grasping, 50% at lifting an object and 27% at pick-and-place.
Our take: This is a strong sign that robot world models may scale like language models: more compute gives better results, and you can predict how much better in advance. Planning from just a goal image is promising, but 27% on pick-and-place shows how far there is to go.
On the community radar
State of AI 2026 meetup, 3 November, Paris
Air Street is presenting the 2026 State of AI Report at an evening meetup in Paris. One fireside chat features Guilherme Penedo of Macrodata Labs on data for training more capable robotics models. RSVPs are open.
Physical AI Builders Day, 2 December, Paris
AI Tinkerers Paris is hosting a one-day physical AI hackathon covering robotics, embedded systems and sim-to-real, supported by apidays, Swanbase and Epitech. You need to apply to attend. The day before, 1 December, the PAIR(ED) Builder Summit covers physical AI, robotics, edge computing and drones.
Help shape the next edition
What are you building, reading or stuck on? Send us your robot videos and photos, papers, open-source projects, articles, job openings or a question for the community.
Send your contribution to [email protected]. We’ll gather, edit and curate submissions for the next digest.
Tell us what you think. This is our first edition, and we want to make it better. Whether you found it useful or not, let us know. Every bit of feedback helps. Just write to [email protected].
Get involved!
Le Physical is a community project, and we need help putting it together. If you’d like to spot news worth sharing, write short summaries, edit, or take photos at our meetups, we’d love to have you on board. Even a couple of hours every two weeks makes a real difference.
Write to [email protected] with “Get involved” in the subject, and tell us a bit about yourself and how you’d like to help.
See you in the next edition,
Miguel & Stan