Embodied R1

A large vision‑language model designed for general robotic manipulation, it bridges the "seeing‑to‑doing" gap by translating visual perception into actionable motor commands. The model incorporates a novel pointing mechanism that localizes target objects and a reinforced fine‑tuning training regime that improves task grounding and zero‑shot generalization. It is trained on a diverse dataset of simulated and real‑world manipulation scenarios, enabling it to handle a wide range of objects and tasks without task‑specific fine‑tuning. The inference pipeline accepts an image and a natural‑language instruction, producing a visual plan that a robot can follow. Researchers and developers in robotics and AI can use the model to prototype manipulation behaviors, accelerate training, and evaluate zero‑shot performance. The project also provides scripts for inference and evaluation, facilitating rapid experimentation.