Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation

TL;DR

This work addresses the seeing-to-doing gap in embodied AI by introducing a unified pointing-based representation and a two-stage Reinforced Fine-tuning curriculum. It presents Embodied-R1, a 3B Vision-Language Model trained on the Embodied-Points-200K dataset with a multi-task reward design to enhance spatial reasoning and embodied pointing. The model achieves state-of-the-art results on 11 embodied-spatial benchmarks and demonstrates strong zero-shot generalization to SIMPLEREnv and eight real-world XArm tasks without task-specific fine-tuning, along with robustness to visual disturbances. Overall, the pointing-centric approach provides a generalizable pathway to bridge perception and action for broad robotic manipulation tasks.

Abstract

Generalization in embodied AI is hindered by the "seeing-to-doing gap," which stems 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 use a wide range of embodied and general visual reasoning datasets as sources to construct a large-scale dataset, Embodied-Points-200K, which supports key embodied pointing capabilities. We then train Embodied-R1 using a two-stage Reinforced Fine-tuning (RFT) curriculum with a 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.

Figures (11)

• Figure 1: The Embodied-R1 framework for zero-shot robotic manipulation through "pointing". Embodied-R1 takes visual and textual instructions, performs explicit reasoning, and then generates a visual trace as a universal command. The other panel showcases our comprehensive evaluation, including spatial reasoning, embodied pointing benchmarks, and real-world robot tasks.
• Figure 2: Overview of four embodied pointing abilities.
• Figure 3: Overview of training data: In stage 1, we focus on improving the model’s spatial reasoning capability, while incorporating a small amount of general reasoning data. In stage 2, we train the model’s embodied pointing capabilities, which comprise four distinct capability items.
• Figure 4: Visualizing Embodied-R1's Performance on Various Pointing Tasks.The model can follow diverse text instructions and generalize its capabilities to novel, unseen environments.
• Figure 5: The process of Embodied-R1 performing real-world tasks.