From YouTube to Robot Arms: The Passive Data Pipeline

Research

Training Robot Manipulation from Passive Internet Data

This paper presents a method for training robot manipulation skills by watching internet videos of humans, not by teleoperating a robot. It's a direct attack on the data bottleneck: instead of needing a robot in the loop for every new skill, you can now bootstrap from passive human demonstrations, which are orders of magnitude easier to collect.

Human Video Understanding Benchmark for Robotic Manipulation

A new benchmark for evaluating how well robots learn from human video shows the gap between watching and doing is still real, but the approaches to close it are converging. If you're building a robot that needs to parse complex object interactions (like opening containers or using tools), this paper is your new baseline for what

Fast Skill Adaptation for Robot Manipulation

I'm not sorry to be skeptical, but this is a solid piece of plumbing work: a framework for quickly teaching a robot a new manipulation skill with a handful of demos. It's not general intelligence, it's a practical adapter that makes existing robot arms more useful without weeks of retraining.

Tools

NVIDIA Isaac Sim Update: Easier Digital Twins

NVIDIA's latest Isaac Sim release focuses on making high-fidelity digital twins easier to build and validate. For anyone doing sim-to-real transfer, this cuts down the plumbing work between your CAD model and your reinforcement learning policy. Less time wrangling meshes, more time testing control.

Events

ROSCon 2025: Production-Ready ROS 2 Updates

ROSCon 2025 is happening, and the pre-announcements are all about production readiness: better security, long-term support for ROS 2 Jazzy, and tooling for managing fleets of robots in the real world. If your stack is ROS, this is where you'll hear what's actually stable enough to ship.