30 minutes of video.
Robot learns the task.
Open-source, end-to-end.
An open-source framework for training robot policies from only 30 minutes of human egocentric videos captured via Meta Aria glasses:
Achieving zero-shot transfer to robots without any robot data collection.
The method relies on Interaction-Centric Tokens that encode hand-object spatial relationships invariant to embodiment and viewpoint, supplemented by auxiliary objectives like object motion prediction and latent consistency to extract richer supervision signals from the same data.
HumanEgo demonstrates strong cross-embodiment, cross-environment performance on bimanual tasks, outperforming baselines like ACT and teleop data while being trainable on a single RTX 4090 GPU.
Thanks for sharing, @TX_Leo_Wang.
📌 Website: https://t.co/i7EDhd7WkX
Paper: https://t.co/0QWfzvRUQ7
Code: https://t.co/0Iry7tr3ep
Video: https://t.co/NSlhRBT1p7
——-
Weekly robotics and AI insights.
Subscribe free: https://t.co/9Nm01QUcw3
If you want to get into training robots badly:
You can start training real robots at home for under $300...
Thread 👇 (with links)
Here’s the exact low-barrier roadmap the community is using right now to go from zero to running ACT policies on a real arm in weeks.
Thread 👇
1/ Simulation First (Zero Cost)
Start in MuJoCo... honestly: the gold-standard physics engine used by DeepMind and every serious robotics lab.
Train policies, test behaviors, and iterate insanely fast before touching hardware.
→ Official site: https://t.co/ou3IXwVseF
→ GitHub (install in minutes): https://t.co/vDwD1B4Jt7
Pro tip: MuJoCo runs perfectly on a normal laptop.
2/ Grab the $200-300 Hardware Everyone’s Using
The SO-101 (also called SO-ARM101): the official open-source arm designed by Hugging Face’s LeRobot team.
3D-printed, easy to assemble, and built specifically for LeRobot + ACT training.
Buy options:
→ WowRobo DIY/Assembled Kit (~$199): https://t.co/X4wJJQovFg
→ Full Hugging Face SO-101 docs + parts list: https://t.co/4zibGFkozt
(Yes, real people are training on this exact arm for a few hundred bucks.)
3/ Install LeRobot + Train ACT Policies
LeRobot is Hugging Face’s open-source robotics library (PyTorch-native).
It ships state-of-the-art policies like ACT (Action Chunking Transformers); perfect for precise manipulation tasks.
→ Main repo (star it): https://t.co/CPKsrrJcDE
→ ACT policy guide + training examples: https://t.co/WWPapXepCN
One-command training examples are in the repo. You’ll be running policies the same day.
4/ Add Multimodal LLMs for Spatial Intelligence
Hook up a cheap webcam + multimodal LLM (like LLaVA, Qwen-VL, or even Claude-3.5) to:
• Understand your workspace in real time
• Select the right ACT policy based on what it sees
• Give natural-language commands
Exactly what @KuphDev is doing in his livestreams (color-sorting ducks with SO-101 + LeRobot + multimodal vision).
This is the “spatially aware” leap everyone’s excited about.
5/ Sim2Real Transfer (The Magic Step)
Train in MuJoCo → fine-tune on the real SO-101.
LeRobot makes domain randomization and sim-to-real dead simple.
Start with basic pick-and-place, then level up to multi-step tasks.
6/ Next-Level Moves (Once You’re Hooked)
• Add a second arm for bimanual tasks
• Train with human demos (teleoperate the leader arm)
• Scale to mobile bases or even low-cost humanoids later
Total starter cost: simulation = free + SO-101 ≈ $200-300 + webcam + laptop you already own.
The barrier is gone.
Hardware access was the only thing stopping most people... now it’s solved.
Who’s actually starting their robot journey this week?
Drop 🔥 if you’re in (or building right now).
Photo: Google Gemini
——-
Weekly robotics and AI insights.
Subscribe free: https://t.co/9Nm01QUcw3
🇧🇷 En Brasil, una madre de 43 años, apuñaló, cortó el pene y luego mató a un hombre que abusó sexualmente de su hija de 11 años.
Tras casi 1 año de estar detenida fue absuelta de manera unánime por legítima defensa de su hija ⚠️🚨
¿Estás de acuerdo? 🙋🏻♀️
@fogondpalo Zapote, hay de dos tonos de colores y formas. A veces puede ser ovalada y otras veces puede de forma eliptica. El Nispero es mas pequeño que el Zapote, tiene un tono de color diferente, es mas pequeño.
@spigen Dont bother me at all the idea for the Royal Pop to use this type of option as an extension about how you can use the same watch for different occasion. My worry would be avoid any scratch when you put the Royal Pop into the band, the material, easy put in and out and so on
I was really hoping to buy a Swatch & AP next weekend, assuming I'd be able to wear it on my wrist. But now that I see it's going to be a pocket watch, I'm having second thoughts...
The Department of War has released the first round of files connected to unidentified aerial phenomena, and conspiracy theorists are sure to have a field day with the docs.
@LagannMikhail I would like to know how you connected the quest 3 with the controller to interact with the movement of the robot arm. Are you using unity for this project or different engine.
Paciencia y genética: la fórmula de Quico Fernández para producir un jamón de bellota en Salta que compite con los grandes de España. 🌳🍖
70 años de árboles y 40 años de selección de raza para llegar a un nivel de altísima calidad internacional. Si están por Salta, pasen por Cerdo Negro y comprueben cómo un productor argentino se puso a la par de los mejores del mundo. Podemos decir que es un Jamón Argentino Tipo Ibérico.
@sparklabhq@arduino Ok,look good. What about to add 4 more colors, increase the speed, add a timer with display and add a screen to see the score of each player. For each correct color pressed player win 10 points. The winner is the player with the most highest score.