正在加载视频...

视频加载失败

learning with just one video🦾

16,797 次观看 • 1 年前 •via X (Twitter)

11 条评论

kfant 的头像
kfant1 年前

with all the gpu compute at the background?

ARX 的头像
ARX1 年前

only one 4060 laptop

Mobile Scanner 的头像
Mobile Scanner1 年前

Scan any documents, convert images into text, PDF files, etc. 👍

Tina Chopra 的头像
Tina Chopra1 年前

The post from ARX states: "learning with just one video" and shows a robotic arm manipulating objects. This implies impressive progress in robotic learning with minimal data. Here's an English comment: "This is incredibly impressive! Learning complex manipulation tasks from just one video is a massive leap forward for robotics. It really highlights the power of advanced AI in accelerating robotic capabilities. What's the most intricate task you envision these ARX robots mastering next with this efficient learning method?

Phil Trubey 的头像
Phil Trubey1 年前

Can you do longer multi step manipulation learning with only one video?

Tensor Templar 的头像
Tensor Templar1 年前

Trying vjepa 2 I suppose?

Luoyang Precision Bearing Co., Ltd 的头像
Luoyang Precision Bearing Co., Ltd1 年前

👍

Ilir Aliu - eu/acc 的头像
Ilir Aliu - eu/acc1 年前

Most robot policies forget what just happened. Makes even simple tasks surprisingly hard. 📍 A new method fixes that... and learns 3× faster. The model needs memory, but a long history usually makes training worse. This paper shows how to fix that with a small trick called Past-Token Prediction: ✅ Adds memory without making the policy worse ✅ Trains faster by reusing earlier features ✅ Works in real-world robot tasks ✅ Lets the robot check if it's remembering correctly during rollout Thank you for sharing, @marceltornev! Full paper, code, and videos here: 📍Paper: Project: Code:

pfung 的头像
pfung1 年前

I'm obsessed w/ robots which can do real-world tasks reliably (vs general intelligence), and created a repo tracking papers there. Please PR if you have more ideas!

Ville 🤖 的头像
Ville 🤖1 年前

My next mission is to build Reliable Robot™ Here's 5 minutes of fully autonomous self-play. It's not perfect yet, as there are a few grasp failures and especially towards the end the robot gets quite shaky (overheating maybe?). But no manual interventions required 😃

Jack 🤖 的头像
Jack 🤖1 年前

Actuators eat up 30-50% of a humanoid robot's BOM cost. Mass adoption? Not at these prices. This is one of the most persistent challenges in robotics, but @IMSystemsNL may have cracked the code with a novel drive. To understand it, we need to dive into the world of drives & transmission, with all their demands and trade-offs. Shoutout to @MarwaEldiwiny and @GoingBallistic5 for the original video.

相关视频