About
Open human–robot and human–agent interaction.
OpenHRI is an open research effort on how machines take part in live conversation with people: when to speak, when to stay quiet, what to do with the face and the voice while the other person is talking, and how to measure any of it.
How we work
- Open by default. Code is released under permissive licenses. Models and dataset manifests go to Hugging Face. Where a source corpus cannot be redistributed, we publish the loader and the recipe instead of the media.
- Numbers you can rerun. Every result in a post names the script, split and settings that produced it, and states its uncertainty.
- Measured in real time. Turn-taking is a timing problem. We evaluate agents by streaming audio and video into them as it happens, not by scoring transcripts after the fact.
Get involved
Issues and pull requests are welcome on any repository at github.com/OpenHRIai. If you use one of the benchmarks or models in a paper, each post has a citation block at the end.
For anything else, write to hello@openhri.ai.