Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation

Yifu Yuan, Haiqin Cui, Yaoting Huang, Yibin Chen, Fei Ni, Zibin Dong, Pengyi Li, Yan Zheng, Jianye Hao
Tianjin University
ArXiv Code Dataset Model

Abstract

Generalization in embodied AI is hindered by the "seeing-to-doing gap", stemming from data scarcity and embodiment heterogeneity. To address this, we pioneer "pointing" as a unified, embodiment-agnostic intermediate representation, defining four core embodied pointing abilities that bridge high-level vision-language comprehension with low-level action primitives. We introduce Embodied-R1, a 3B Vision-Language Model (VLM) specifically designed for embodied reasoning and pointing. We train Embodied-R1 using a two-stage Reinforced Fine-tuning (RFT) curriculum with specialized multi-task reward design. Embodied-R1 achieves state-of-the-art performance on 11 embodied spatial and pointing benchmarks. Critically, it demonstrates robust zero-shot generalization by achieving a 56.2% success rate in the SIMPLEREnv and 87.5% across 8 real-world XArm tasks without any task-specific fine-tuning, representing a 62% improvement over strong baselines. Furthermore, the model exhibits high robustness against diverse visual disturbances. Our work shows that a pointing-centric representation, combined with an RFT training paradigm, offers an effective and generalizable pathway to closing the perception-action gap in robotics.

Embodied-R1 Framework
Overview of Embodied-R1 Framework. Embodied-R1 is a 3B vision-language model (VLM) designed for general robotic manipulation. Through an innovative "Pointing" mechanism and Reinforced Fine-tuning (RFT) training methodology, it effectively bridges the "seeing-to-doing" gap in robotics, achieving remarkable zero-shot generalization capabilities.


Results

Real-world Robot Manipulation Demonstrations

Embodied-R1 demonstrates superior performance in real-world robot manipulation tasks, achieving an 87.5% success rate across 8 diverse tasks in a zero-shot setting. Below are visualizations of the model performing these tasks.

Real-world experimental evaluation results
Real-world experimental evaluation results.

Task 1: Pick up the strawberry

Task 2: Move the egg to the bowl

Task 3: Move the vise to the red basket

Task 4: Place the fork in the green bin

Task 5: Pick the [x] toothbrush and place it to the bucket

Task 6: Move the nearest object to the right side of the drawer

Task 7: Put the screwdriver between drawer and the vase

Task 8: Move the moka pot to the right of drawer



Visualization of Performance on Various Pointing Tasks

Embodied-R1 Visualization
Visualizing Embodied-R1's Performance on Various Pointing Tasks. The model can follow diverse text instructions and generalize its capabilities to novel, unseen environments.

Additional Embodied-R1 Results


Robustness to Visual Disturbances

Visual Disturbance Results
Embodied-R1's performance under different visual disturbance conditions. The model maintains 100% success rate under original conditions and background changes, with 83% success rate under combined background, light, and height changes.

Original

Background Change

Background+Light Change

Background+Light+Height Change

The process of Embodied-R1 performing Task 6 under different visual disturbances.

BibTeX

@misc{yuan2025embodiedr1reinforcedembodiedreasoning,
            title={Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation}, 
            author={Yifu Yuan and Haiqin Cui and Yaoting Huang and Yibin Chen and Fei Ni and Zibin Dong and Pengyi Li and Yan Zheng and Jianye Hao},
            year={2025},
            eprint={2508.13998},
            archivePrefix={arXiv},
            primaryClass={cs.RO},
            url={https://arxiv.org/abs/2508.13998}, 
      }