Ran molmoact2 zero shot on SO101! And it is my first policy on the robot.
I added support on Lerobot to spin up molmoact2 on a VM since I did not have the capability to run the model on my client machine.
https://t.co/gNRG8iBIWE
@allen_ai@DJiafei@hq_fang
Diffusion models are easier to understand when you can see the geometry move.
Diffusion Explorer is an interactive educational tool for understanding the geometric intuition behind diffusion and flow-based generative models.
It helps you connect equations to model behavior by visualizing generated samples over time, showing how samples change through training, and letting you train on custom hand-drawn distributions.
Key features:
• Flow + score objectives – implements training objectives like Flow Matching and Denoising Score Matching
• Sample dynamics view – shows how generated samples evolve over time for pretrained models
• Training-time visualization – lets you observe how generated samples change as training progresses
• Custom distributions – supports training on hand-drawn distributions for faster intuition building
• Rectified Flow explainer – includes animated visuals for curved trajectories, few-step sampling issues, straightening, and optimal transport links
Free public GitHub repo.
Link in the reply 👇
The hand is the missing interface layer between AI and the physical world.
Worth noting where the components are actually being built.
The dexterous hand ecosystem, both the components and integrated products, is consolidating across a handful of Asian suppliers.
✋🏻 Finished hands:
- @SharpaRobotics Wave, @wuji_global, @PaXiniTech Tech's DexH13, @AGIBOTofficial spinout AgiLink’s OmniHand, and XYNOVA's Flex 2.
Just a few examples that are shipping commercially.
⚙️ The component tier:
- micro harmonic reducers (Leaderdrive, Laifu Harmonic),
- precision ball screws (Wuzhou Xinchun),
- embedded controllers and flex PCBs (Leadshine, Avary, Dongshan).
In terms of sensors, @gelsight is developing a miniaturized "digital fingertip" and soft tactile sensor with a Phase II Small Business Innovation Research (SBIR) contract from the United States Air Force.
XELA, a Tokyo/Waseda spinout, is unveiling several new major tactile sensor enhancements at Automate in Chicago later this month.
The dexterous hand is not just a hardware business.
It’s a data business.
Whoever ships the most tactile sensors and dexterous hands, accumulates the most tactile data.
And that is the key missing piece in training better models.
Train AI robots without writing a single line of code. 🤖
We just launched LeLab, the official graphical user interface for LeRobot built by @rabault_nicolas. It completely removes the command line from the robot learning workflow, taking you from raw hardware to autonomous movement visually.
If you've ever wanted to get into AI robotics but were held back by complex terminal setups, this is for you.
- Zero-Terminal Setup: Smart calibration with automatic USB port detection.
- Easy Data Collection: Teleoperate your robot and record a dataset.
- One-Click GPU Training: Don't have a massive local GPU? Scale your training instantly with Hugging Face Jobs right inside the app.
Just plug in your SO-ARM101 and start teaching your robot. We put together a complete, step-by-step video guide showing exactly how to get started and train your first policy.
Docs: https://t.co/PrUEIeaXKW
GitHub: https://t.co/SFuOiN8rjN
Ran molmoact2 zero shot on SO101! And it is my first policy on the robot.
I added support on Lerobot to spin up molmoact2 on a VM since I did not have the capability to run the model on my client machine.
https://t.co/gNRG8iBIWE
@allen_ai@DJiafei@hq_fang
Simulate an Unmanned Aerial Vehicle (UAV) using camera and lidar sensor models to deliver a package while navigating a city environment and avoiding obstacles 🛫
🔗 https://t.co/jfuuaD5e6n
Simulate an Unmanned Aerial Vehicle (UAV) using camera and lidar sensor models to deliver a package while navigating a city environment and avoiding obstacles 🛫
🔗 https://t.co/jfuuaD5e6n
• ROS is free
• Gazebo is free
• Isaac Sim is free
• RViz is free
• MoveIt is free
• PX4 and ArduPilot are free
• OpenCV is free
• PCL is free
• TensorFlow and PyTorch are free
• KiCad is free
• Fusion360 is free for hobbyists
• Linux is free
a laptop, an internet connection, and an obsession with robots; that’s enough to start building machines that move in the real world.
NeurIPS 2025's top 50 paper contributors show China and the US neck-and-neck.
In the US, corporate labs—Google DeepMind, Meta, Microsoft—now rival top schools like Stanford, CMU, MIT, showing a major shift to industry-led AI research.
China’s leaders remain academic: Tsinghua, CAS, Peking, SJTU, HKUST. Singapore (NUS, NTU), Korea (KAIST), UAE (MBZUAI), and Canada (Mila) hold strong.
Europe trails, with only Oxford, EPFL, ETH Zurich, and TUM from the EU-27 making the list.
---
Source:
linkedin. com/posts/pierre-alexandre-balland-20b75b13_who-pushed-the-ai-frontier-at-neurips-2025-activity-7403119036496162817-vRPE
Ran my blog post about FoloUp through @NotebookLM , and the new @NanoBanana templates are ridiculously good! It produced a video that’s genuinely excellent, and the narrative it created made the story ~10x stronger than my own draft.
Best places to travel in October, 2025.
1. 🇱🇰 Sri Lanka
2. 🇹🇷 Turkey
3. 🇺🇸 New Mexico
4. 🇪🇸 Valencia
5. 🇺🇸 New York City
6. 🇵🇭 Philippines
7. 🇧🇹 Bhutan
8. 🇧🇦 Bosnia and Herzegovina
9. 🇷🇴 Timisoara, Romania
10. 🇺🇸 San Francisco
11. 🇳🇦 Namibia
(Timeout)