RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance

Dongchi Huang; Hongyin Zhang; Bohan Hou; Siteng Huang; Zhian Su; Hang Guo; Tong Lu; Zhaofeng Xu; Jiahao Tang; Jianfei Yang; Donglin Wang; Peixi Peng; Mingxiu Chen; Deli Zhao; Xin Li

RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance

A general-purpose value model that understands how far a robot is from completing an instruction—and turns that signal into progress estimation, failure detection, and rewards for real-world learning.

Dongchi Huang^* Hongyin Zhang^* Bohan Hou^* Siteng Huang^† Zhian Su Hang Guo Tong Lu Zhaofeng Xu Jiahao Tang Jianfei Yang Donglin Wang Peixi Peng^† Mingxiu Chen Deli Zhao Xin Li

DAMO Academy, Alibaba Group Hupan Lab * Equal contribution · † Corresponding authors

📄 Paper 🐙 GitHub 🤗 HuggingFace 🪐 ModelScope

RynnValue learns value through temporal distance across diverse robot embodiments and viewpoints.

01 · Method

Value, grounded in time.

A multimodal backbone predicts absolute and relative value while a language head verifies instruction alignment. Value-isolation attention keeps temporal supervision focused across long clips.

RynnValue architecture, heterogeneous training data, value heads, and downstream robot tasks — RynnValue architecture—from heterogeneous robot data to progress, failure, and reward signals.

RynnValue temporal sampling, shuffling, instruction mismatch augmentation, and value-isolation attention — Training recipe: temporal sampling, order shuffling, instruction mismatch augmentation, and value-isolation attention.

02 · Trajectory-ranking results

Generalizes beyond the training distribution.

RynnValue is evaluated on six out-of-distribution robot datasets. The 8B model achieves the strongest average trajectory-ranking result, while the ablation isolates the contribution of each design component.

RynnValue trajectory-ranking results on RBM-EVAL-OOD — Trajectory ranking across six out-of-distribution robot datasets.

RynnValue ablation results for shuffling, isolation, language supervision, and random sampling — Component study on temporal-order shuffling, isolation, language supervision, random temporal sampling, and relative modeling.

03 · Reinforcement learning results

RynnValue as a real-world reward.

Online and offline robot learning consistently benefit from the richer value signal, with RynnValue highlighted in purple.