OpenGalaxea/GalaxeaVLA

What it solves

GalaxeaVLA provides a general-purpose robot control system that enables robots to perform complex manipulation tasks by combining visual perception, natural language instructions, and precise motor actions. It addresses the challenge of bridging the gap between high-level reasoning (understanding what to do) and low-level execution (how to move the robot arm) within a single model.

How it works

At its core is G0.5, an autoregressive vision-language-action (VLA) model. Unlike traditional systems that use a VLM as a separate encoder, G0.5 uses a single transformer decoder to generate reasoning tokens and action tokens in one continuous stream.

Key technical components include:

• Unified Stream: The model processes multi-view RGB images, robot state, and text instructions to output a sequence that first includes embodied reasoning (like subtasks and object grounding) and then structured action tokens.
• Cross-embodiment ActionCodec: A learned tokenizer that maps diverse robot actions into a shared 27-dimensional action space, allowing the model to generalize across different robot hardware.
• Native Chain-of-Thought: Reasoning is integrated directly into the token sequence, meaning the model "thinks" about the task (e.g., identifying a bounding box for an object) before emitting the actual movement commands.
• Visual Memory: The model uses factorized spatial-temporal attention to incorporate several seconds of visual history, improving stability and context awareness.

Who it’s for

This project is designed for robotics researchers and developers working on embodied AI, specifically those deploying models to real-world robots (such as R1 Lite, R1 Pro, SO-100/101, and Franka) or testing in simulators like LIBERO and RoboTwin.

Highlights

• Single-Stream Architecture: Combines reasoning and action into one next-token prediction objective.
• Broad Hardware Support: Includes deployment entrypoints for multiple real-world robot embodiments.
• High Performance: Demonstrates strong zero-shot success rates on benchmarks like DROID (82.5%) and LIBERO (98.9%).
• Open-World Dataset: Provides over 500 hours of real-world mobile manipulation data in RLDS and LeRobot formats.