Ivan Prusov’s Post

Google DeepMind’s AI System Teaches Robot to Perform Complex Task. Here's what's interesting: These robots have learned from human demonstrations. And they have translated images into action. What you're seeing is two robot arms coordinating and completing tasks that have historically been very difficult for robots to perform, and they're performing them autonomously. This is a breakthrough in robotics, thanks to new AI systems. Really makes you wonder where robotics is going next.

New innovation

🚀 Universal Manipulation Interface (UMI) is revolutionizing the way robots learn and perform tasks. 🛠️ An innovative framework that facilitates seamless transfer of skills from human demonstrations to robot policies developed by researchers from Stanford University, Columbia University and Toyota Research Institute. UMI employs hand-held grippers coupled with careful interface design to enable portable, low-cost, and information-rich data collection for challenging bimanual and dynamic manipulation demonstrations. 🤖 With its focus on portable, low-cost data collection and hardware-agnostic policy learning, UMI opens doors to dynamic, bimanual, precise, and long-horizon behaviors across various robot platforms. #Robotics #Innovation #UMI #AI

Integrating UMI into a business network: 👉 https://t.me/theunityhub can empower companies to leverage robotic technology more effectively, driving efficiency, scalability, safety, and innovation across various industries and applications.

NVIDIA unveils Project GROOT, teaching humanoid robots to navigate both the real and digital realms. 🤖 GROOT learns from videos and example demonstrations, just like we do from YouTube tutorials. It processes these inputs to perform the next move. Similar to how kids learn from watching and listening, you could show it how to high five a couple of times and then it can high five someone on its own. That is the future NVIDIA is building. 🚀 What's your take on this development? Drop your thoughts below! #NVIDIA #Robotics #AIInnovation #ProjectGROOT #AI #Tech #Innovation #TechFun

Can you remind me again why we don't want robots taking over our roles? Let me reiterate: we definitely don't want robots handling any tasks... Should we really stop teaching them new skills?! Alright, kidding aside... Imagine a way to teach a robot a new skill without even needing a physical robot! That's exactly what UMI does. It uses a special hand gripper equipped with a camera to learn from demonstrations by humans, without requiring an actual robot. Moreover, UMI can teach the same skill to different robots, even if they are constructed differently. This advancement means robots can learn intricate tasks like grasping objects or operating tools, significantly enhancing their usefulness in practical scenarios. I'm sold. I want this robot right now! Stay tuned for more updates on technology. #artificialintelligence #emergingtechnology #robotics

Mobile ALOHA: The Budget-Friendly Robot Chef from Stanford University Stanford's Mobile ALOHA is shaking things up in the world of robotics! This low-cost, open-source robot learns through human demonstrations, tackling complex tasks like cooking and navigating environments. Imagine a future with an AI-powered housemaid - Mobile ALOHA paves the way! #StanfordRobotics #MobileALOHARobot #AIHousemaid #FutureofWork #RobotChef #OpenSourceHardware #BimanualManipulation #ai #artificialintelligence #agi #artificialgeneralintelligence #machinelearning #deeplearning #job #vacancies #robots #robotics #software #developers #urgentvacanies #google #microsoft #cognizant #wipro #tcs #tata

NVIDIA recently announced something pretty groundbreaking: Project GR00T and a big update to their Isaac Robotics Platform. This is a big one. Picture robots that not only understand and speak your language but also learn from watching your moves. That is the promise of Project GR00T. It feels like we successfully entered a new era of robots. Just last week, I saw a demo of GPT-integrated Humanoid from a Robotics company Figure: https://lnkd.in/detaC8yB. Nvidia also introduced Jetson Thor, a powerhouse of a computer designed specifically for these advanced robots, featuring top-notch AI performance capabilities. Plus, the Isaac platform got a serious boost, including new simulation tools and AI workflows that could dramatically change how robots learn and operate. Imagine having a robot in your life that learns to help with daily tasks just by observing. Cooking, playing, cleaning so on and so forth. Share this to spread the word about the future of robotics.

In the not so distant future 1 billion robots will live and work among us and they won't need centralised GPUs or an Internet connection to operate. All visual, NLP and Audio will be processed locally on neural accelerator or GPU. The biggest breakthroughs of the next 5 years will not be larger LLMs, but rather inference frameworks that let you achieve more with less GPU. For the last 12 months me and my team at YEPIC AI have been building the world's smallest Multimodal video model for generating emotional intelligent interactive Avatars. Before robots live among us, the next logical step is multimodal face to face interaction with AI.

How can we efficiently train robots to solve tasks with sparse rewards using only image-based observations? Check out our paper "LaNE: Accelerating Visual Sparse-Reward Learning with Latent Nearest-Demonstration-Guided Explorations" at #CoRL2024! #reinforcementlearning #robotics We introduces a sample-efficient approach to tackle sparse-reward robotic manipulation from high-dimensional pixel data, using just a few demonstrations and deep reinforcement learning. Our method LaNE leverages a pre-trained DINOv2 feature extractor to create an embedding space with locally-linear forward dynamics. We derive an exploration bonus from the embedding space to guide the agent to explore environments effectively near the demos. LaNE improves exploration without relying on dense, hand-crafted reward functions. Experiments showcase LaNE's capability to achieve state-of-the-art sample efficiency in Robosuite simulations and learn tasks in under an hour on a real Franka Panda robot! 📌 Project Page: https://lnkd.in/gerAHCwS 📌 Paper: https://lnkd.in/gYAgppKT This work was an exciting collaboration with Ufuk Topcu, Sandeep Chinchali and Mariano Phielipp.