REKI couples a vision-language model, which encodes camera views and instructions, to a flow-matching action expert that generates chunks of robot actions.
Compact robot policy uses lightweight memory to stabilize action generation
REKI combines public-data pretraining with cached register tokens that preserve recent visual context at less than 1 percent inference overhead.
Research Lab
Juliette Marrie · Paul Pacaud · Adrien Ramanana Rahary · Shizhe Chen · Moritz Böhle
Kyutai · Inria · École nationale des ponts et chaussées
Research Digest··2 min read
Thread:Robot Manipulation Policies
4-billion-parameter vision-language-action model for language-conditioned robot control.
Why this paper
From Inria and 2 others · Part of Robot Manipulation Policies, now 25 papers
In one line
A 2.4B vision-language-action model trained only on public data is outperformed only by models about twice its size, and a register-token memory cuts pose oscillations fourfold.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (hardware H100)
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§