Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models
Yifu Yuan, Yaoting Huang, Xianze Yao, Yutong Li, Shuoheng Zhang, Linqi Han, Pengyi Li, Jiangeng Sun, Wenting Jia, Zhao Zhang, Yuhao Liu, Ruihao Liao, Yucheng Hu, Qiyu Wu, Yuxiao Li, Zibin Dong, Fei Ni, Yan Zheng, Shuyang Gu, Yi Ma, Hongyao Tang, Han Hu, Jianye Hao
arXiv:2606.11324 · 170▲ · cs.RO, cs.AI, cs.LG
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
Embodied-R1.5 is an 8B-parameter vision-language model fine-tuned to unify three robotics-relevant capabilities (spatial reasoning, task planning/correction, and pointing/trajectory generation) in one model, trained via supervised fine-tuning followed by reinforcement learning with verifiable rewards on 15B tokens of mostly auto-generated data. It also ships a Planner-Grounder-Corrector loop where a single model instance plays all three roles for long-horizon tasks, and can be extended into a robot action policy (VLA) by attaching a flow-matching action head.
Why it matters
The paper's central practical claim is that if a VLM has already internalized embodied reasoning, you can turn it into a working manipulation policy with only a small amount of action data instead of large-scale action pretraining, which would cut the data collection cost that dominates robot learning. All outputs are plain-text tokens (coordinates normalized to 0-1000), so there is no custom vocabulary or special-token machinery to integrate.
Practical takeaway
Model weights, datasets, training code, and an evaluation harness (EmbodiedEvalKit, covering 25+ embodied benchmarks) are open-sourced, so you can run standardized embodied-VLM comparisons yourself instead of writing ad-hoc parsing scripts. Watch for the pattern of using RL with verifiable rewards (IoU, point distance, trajectory RMSE) to specialize a general VLM for grounding tasks.
Key result
Reported state-of-the-art on 16 of 24 embodied VLM benchmarks with a 70.4% average across 21 accuracy benchmarks, versus 53.4% for Gemini-Robotics-ER-1.5. As a VLA it reports 92.4% on SimplerEnv Google Robot Visual Matching (over 20 points above pi-0.5). Caveat: results are self-reported through the authors' own evaluation framework, and the real-robot experiments are qualitative demonstrations rather than quantified success rates.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.