Open physical-AI skills: from human demonstrations toward installable robot skills.
Official site · Research notebook · Hugging Face
Train locally from Skillspace footage on your computer or server. The open-source pipeline checks videos, keeps related recordings together, fits small CPU models, and exports inspectable results. Training runs offline after the software and data are installed; this learner requires no GPU.
Install locally · Hardware guide · Recorded demo
Source code · Synthetic datasets · Pilot models · Frozen benchmark evidence
Research alpha. Visual prediction and robot movement are separate claims. Movement training requires aligned numerical actions and a supported simulation adapter. The demo presents the original one-scene recorded pilot; the separate sample-count benchmark tests installed candidates in twelve simulated layouts. Human-video action recovery and physical-arm qualification remain open research tasks.
We publish methods, reproducible inputs, measured outcomes, controls, and failures in the research notebook. Start with the footage-training pipeline and frozen sample-count follow-up, or inspect the research source and revisions.
To contribute, see the training contribution guide or the research contribution guide.
Open-weight inference and hardware research live with our sister product, OW Terminal.