RoboHarness: A Simple Harness Outperforms VLA and World Actions Models

Researchers from Kyutai, Inria, and Ecole des Ponts have introduced REKI, a 2.4B parameter vision-language-action model trained exclusively on public data. The model utilizes a compact memory of recent actions to improve performance in robot manipulation tasks. It demonstrates a new balance between model size and capability by outperforming existing vision-language-action models on three out-of-distribution benchmarks.