EgoSteer open-sources dexterous VLA trained on 9.6K hours

The gap that keeps generalist robot policies from doing anything genuinely useful with a dexterous hand has been staring the field in the face for a while: not enough language-aligned, action-accurate demonstration data at the scale a modern vision-language-action model actually wants to eat. A new paper on arXiv, EgoSteer, tries to close that gap in the open.

The pitch is a full-stack release, not just a model checkpoint. There is EgoSmith, a data pipeline that curates in-the-wild egocentric video into what the authors describe as 9.6K hours of high-quality pre-training data, at 9x higher throughput and better accuracy than the prior state of the art. There is a unified robot stack for teleoperation and human-in-the-loop correction. And there is EgoSteer itself, a world-model-enhanced VLA trained on that infrastructure and refined with DAgger on real robots. All of it, the abstract says, is being open-sourced at the project site.

The headline results are the ones you would want out of a steerability paper. EgoSteer, the authors report, robustly executes free-form instructions across 40+ diverse tasks and shows failure recovery, dexterity, and generalization. The more attention-grabbing number is that the pre-trained model few-shot adapts to complex long-horizon tasks, including box folding, on two embodiments with 75+% success.

This is an abstract, not an independent evaluation. The paper as released does not disclose compute budget, latency, or the exact hardware, and it does not put EgoSteer up head-to-head against the well-known closed dexterous VLAs from industry labs. Treat the numbers as author-reported claims until someone else replicates them.

For anyone trying to build dexterous-hand policies without an industrial-scale data budget, the interesting thing is that a reference stack now exists in public, at a scale worth beating.