Northstar

@KyleVedder We’ve been internally validating the data by fine-tuning policies including Pi 0.5 and MolmoAct2 on subsets of our egocentric + task-specific datasets. More on this soon!

@NorthstarRobots what policies did you train with your data? I can’t find any info about that on your website

Most robotics data is slop. That's the real bottleneck for physical AI. Bad angles. No diversity. No quality control. Impossible to train on. We raised $500K from Genius Ventures to fix this. Introducing Northstar.

@jonnbooth @kstonestudios 🧑‍🍳

@NorthstarRobots one of the better launch vids i've seen in a minute

Everything is data. But not all data is trainable. We're building the most high quality & trainable robotics dataset in the world. If you're working on robotics, request a sample today! We're also hiring: ML & software engineers (SF) and Operations (India, Brazil), email us contact@northstarrobotics.ai northstarrobotics.ai

But we don't just collect data. We make it trainable. Our pipeline handles filtering, segmentation, hand tracking, depth maps, multimodal sync, and human + auto annotations- so robotics teams don't have to. We've already validated performance improvements deploying our data on real robots. Labs get training-ready data out of the box.

Real-world egocentric data 🥒