green-vla-robotics

name	green-vla-robotics
title	Green-VLA: Staged Vision-Language-Action Model for Generalist Robots
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.00919
keywords	["Vision-Language-Action","Robot Control","Staged Training","Transfer Learning","Action Spaces"]
description	Train robot controllers via five-stage curriculum progressing from base vision-language models to embodiment-specific RL-refined policies. Unified action space enables cross-embodiment transfer with minimal performance loss.

Green-VLA: Staged Robot Learning

Problem

Robot learning typically requires task-specific pretraining and cannot leverage internet-scale vision-language knowledge. Each embodiment requires separate training from scratch.

Cross-embodiment transfer is difficult because robot actions are heterogeneous—different robots have different action dimensions and semantics. Naive action padding creates spurious penalties from unused dimensions.

Core Concept

Green-VLA implements five training stages (L0→L1→R0→R1→R2) that progressively specialize vision-language models to robot control. The key innovation is a unified action space with fixed semantic layout where action index ranges have consistent physical meaning across robots.

This masked training approach eliminates spurious penalties and enables positive transfer between different embodiments.

Architecture Overview

• L0: Base VLM: Start with large vision-language model (foundational knowledge)
• L1: Web Pretraining: Exposure to internet-scale video and multimodal data
• R0: General Robotics Pretraining: 3000+ hours diverse robot demonstrations
• R1: Embodiment-Specific Fine-Tuning: Supervised fine-tuning on target robot data
• R2: RL Alignment: Reinforcement learning for long-horizon task improvement
• Unified Action Space: Semantic layout with consistent meaning across embodiments

Implementation

Step 1: Build Unified Action Space

Design action space with fixed semantic regions across different robots.

class UnifiedActionSpace:
    def __init__(self):
        # Fixed semantic layout for action indices
        self.action_regions = {
            'base_xy': (0, 2),           # XY movement
            'base_rotation': (2, 3),     # Rotation
            'shoulder': (3, 6),          
            : (, ),             
            : (, ),            
            : (, )          
        }
        .total_dim = 

     ():
        
        mask = np.zeros(.total_dim, dtype=)

         embodiment_type == :
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 

         embodiment_type == :
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 
            mask[.action_regions[]] = 

         mask