The Year in AI Papers
The year/Independent research

Paper 2602.00919

Green-VLA: Staged Vision-Language-Action Model for Generalist Robots

I. Apanasevich, M. Artemyev, R. Babakyan, P. Fedotova, D. Grankin, E. Kupryashin, A. Misailidi, D. Nerus, A. Nutalapati, G. Sidorov, I. Efremov, M. Gerasyov, et al.

Published
Jan 2026
Research lab
Independent
Citations
4
GitHub
138 stars

01 In brief

Summary

Green-VLA is a staged Vision-Language-Action (VLA) framework for real-world robot deployment, developed by Sber Robotics Center.

It uses a five-stage curriculum: L0 (base VLM), L1 (web pretraining), R0 (multi-embodiment robotics pretraining), R1 (embodiment-specific fine-tuning), and R2 (RL alignment).

The framework unifies 24M web samples and 3,000 hours of demonstrations into a 64-dimensional action space with embodiment prompts, using a DataQA pipeline to filter data based on jitter, sharpness, diversity, and state variance.

It also includes a joint prediction module (JPM) for precise object targeting and an out-of-distribution detector.

Green-VLA is optimized for the Green humanoid robot (32 DoF upper body) but generalizes zero-shot to other embodiments.

Experiments show it outperforms baselines like π0 and GR00T N1 on table-cleaning tasks, achieves state-of-the-art results on Simpler benchmarks after RL alignment, and improves long-horizon success rates on CALVIN.

The model is a ~5B-parameter VLA using Qwen3-VL-4B-Instruct as backbone, trained on 64 H100 GPUs.

Future work includes multilingual support and lightweight reasoning modules.

02 From the paper

Abstract

We introduce Green-VLA, a staged Vision-Language-Action (VLA) framework for real-world deployment on the Green humanoid robot while maintaining generalization across diverse embodiments. Green-VLA follows a five stage curriculum: (L0) foundational VLMs, (L1) multimodal grounding, (R0) multi-embodiment pretraining, (R1) embodiment-specific adaptation, and (R2) reinforcement-learning (RL) policy alignment. We couple a scalable data-processing pipeline (3,000 hours of demonstrations) with temporal alignment and quality filtering, and use a unified, embodiment-aware action interface enabling a single policy to control humanoids, mobile manipulators, and fixed-base arms. At inference, the VLA controller is enhanced with episode-progress prediction, out-of-distribution detection, and joint-prediction-based guidance to improve safety and precise target selection. Experiments on Simpler BRIDGE WidowX and CALVIN ABC-D, as well as real-robot evaluations, show strong generalization and performance gains from RL alignment in success rate, robustness, and long-horizon efficiency.