What would robotics look like if everything worked perfectly? The trick is to work backward from that vision.

Here is what goes wrong with these tests !!! (Most common problems ) The models learning to recognizes the tests itself. For example, Sonnet 4.5 can tell when it's being evaluated. You can't tell if they're actually good or just performing. There is a real quote from

The feedback signal in RL is just so weak !! We do all this work (like a min long robot action) but get only one number at the end (the reward). We then try to push that tiny bit of feedback back through all the steps we took, it’s so wasteful and indirect. If you think about it

1. Research 🎯 Goal: Discover new principles or models 📦 Output: Papers, prototypes, foundational ideas ⚠️ Risk: High, but creates the future “No real users yet, just an intuition and a theory & may be a POC.” ——- 2. Technology Development 🎯 Goal: Make research usable at

🤖 Why can’t robots see and manipulate yet? (🔔Just a Reminder) Because physical difficulty ≠ robotic difficulty. What feels easy to humans is still insanely hard for machines. It still requires: ✅ High-quality 3D perception in cluttered, dynamic environments ✅

I get asked many times what is the best way to learn CUDA ? If you are super new to CUDA then you can follow NVIDIA's official learning path 👨🏽‍💻 NVIDIA: https://t.co/PP8Y1akAzt There are books on CUDA but I personally find them little outdated (still worth reading but can be

Concepts > syntax. Welcome to the polyglot era. AI is killing ‘I’m a Python dev’ vibes. The identity of being tied to one language (like “Python developer”) is becoming outdated. Now you’re just a “developer.” LLMs handle syntax, you handle concepts. Your job is to understand

Everyone's talking about AI breakthroughs, but here's why we haven't seen an "OpenAI moment" for robotics yet and what needs to happen before we do. While there's incredible progress happening, we're still missing some critical pieces: The Physical World is Hard AI may be

☕️ Interesting Sunday morning’s ☀️ read. This paper introduces UltraMem, a new way to make big AI models faster and more efficient. As AI models get bigger, they need more computing power and memory, which makes them slow and expensive to run. A popular method called Mixture of

Are Generalist AI Models the Future of Robot Learning? Reinforcement learning (RL) has led to impressive robot demos, but deploying these controllers in the real world remains challenging. Training a policy in simulation is one thing, making it work reliably on a physical robot

Build → Measure → Learn There are 52 weeks in a year. ->> You might lose 1-2 weeks to being sick. ->> Another 1-2 weeks might be bad days/unproductive. ->> Add 2-3 weeks for holidays and breaks. That still leaves you with ~45 weeks to build and experiment. In today’s world,

The future of robotics is open! Excited to see Pi0 by @physical_int being the first foundational robotics model to be open-sourced on @huggingface @LeRobotHF. You can now fine-tune it on your own dataset. 🦾🦾🦾

Handy Notebook: https://t.co/k9dd3xJVgt

Let’s keep the politics aside for the time being and let’s just appreciate the Key technical Innovations that DeepSeek-R1 and DeepSeek-V3 brought into the AI world which helped them compete with and even surpass ( in some cases ) OpenAI’s Models 🚀 . No matter what’s your view

🚀How Cold-Start Data Made DeepSeek-R1 Better? It seems to be very clear from the paper that cold-start data played a big role in improving DeepSeek-R1 by addressing the weaknesses of its predecessor, (DeepSeek-R1-Zero), and boosting its reasoning abilities. 🤔 But first, what

How DeepSeek used distillation to make Smaller Models Smarter? Lately, there’s been a lot of talk about DeepSeek’s innovative methods. One specific approach standout is how they used distillation to make smaller AI models just as smart as the big ones. Big models are powerful,

🚀 NVIDIA’s Cosmos: A New Era for Robotics Training and Simulation Robotics is entering a new era, thanks to NVIDIA's Cosmos World Foundation Model Platform. It is a game-changing tool that makes training robots faster, safer, and smarter. Designed for robotics and other

I think people often believe intuition is some kind of magic.But really, it comes from learning a lot, spotting patterns, and using past experiences to make quick decisions. It’s not about guessing, it’s about figuring out what’s most likely right based on what you already know.