A 22-author paper posted to arXiv on October 26 introduces RynnUnify, a 16-billion-parameter embodied-AI model that folds video-based scene understanding, physical prediction, and robot control into a single system. On the paper’s own long-horizon planning benchmark, Rynn-VP, the model scores 9.75 points higher than Nvidia’s Cosmos 3 Nano, and on a real robot it completes assigned tasks 46.7% of the time on average across three test setups.
By the numbers
The paper’s authors describe RynnUnify as an “interleaved video-language” architecture: rather than treating perception, planning, and control as separate modules, the model reasons in video frames it generates internally before committing to an action. Alongside the base model, the team built an agentic extension trained with reinforcement learning that selectively calls on those internal video predictions to reason about states it cannot directly observe — a technique the paper credits with a 6.0-point average gain over direct reasoning on its companion benchmark, Rynn-GEU.
Pushed into actual robot control through a third extension, the model posts a 51.5% success rate on 20 selected composite tasks in RoboCasa365, a simulated-kitchen benchmark, before dropping to 46.7% on physical hardware — a gap that is typical for sim-to-real transfer in manipulation research but still leaves more than half of attempted tasks failing outright.
The arXiv listing discloses no institutional affiliation for the authors. The “Rynn” name is the same one Alibaba attached to RynnBrain, an open-sourced embodied-AI model the company released earlier this year with claims of outperforming comparable systems from Google and Nvidia on internal evaluations. Whether RynnUnify is a direct successor from the same team could not be confirmed from the paper itself, and this report relies solely on the arXiv preprint — it has not yet been peer reviewed, and no code or model weights have been published alongside it.
Leave a Reply