Start here · Understand the space · Updated 2026-08-23
Open robotics runs on contributions, and the useful ones are not what newcomers expect. Nobody needs your first pull request to the training code. These four lanes create real value from week one.
Community datasets train real models now; SmolVLA pretrained on 481 of them. One well-made dataset, real task labels, consistent cameras, a proper license, is a genuine contribution with your name on it. The recording guide and sharing guide get you there.
A third of popular datasets fail basic checks. Fixing your own before upload raises the floor (pip install datum-lint). Reporting real problems in others, kindly, raises it further. If a Datum check ever gets something wrong, dispute it: disputes are public and change the rubric for everyone.
The field's honest secret is that almost all model scores come from simulation, and everyone knows simulation flatters. Real-hardware evaluation needs a network of ordinary people with standard arms running standard tests. If you finish the ladder and own an SO-101, you are qualified: raise a hand at The Proving Ground.
Every confusing thing you figure out is a contribution waiting to be written. The beginner who writes "here is what actually worked" helps more people than most papers. Post it anywhere public; plain English is the scarcest resource in robotics.
Show your work, be honest about failures, and make things others can build on. Reputation in open robotics compounds exactly like data does. Start where you are on the ladder, and welcome to the space.