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

Authors: I. Apanasevich, M. Artemyev, R. Babakyan, P. Fedotova, D. Grankin, E. Kupryashin, A. Misailidi, D. Nerus, A. Nutalapati, G. Sidorov, I. Efremov, M. Gerasyov, D. Pikurov, Y. Senchenko, S. Davidenko, D. Kulikov, M. Sultankin, K. Askarbek, O. Shamanin, D. Statovoy, E. Zalyaev, I. Zorin, A. Letkin, E. Rusakov, A. Silchenko, V. Vorobyov, S. Sobolnikov, A. Postnikov · arXiv 2026

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

One ~4B VLA trained in five stages (web data, multi-robot pretrain, robot tuning, RL) controls a humanoid, dual-arm and single-arm robots through a shared 64-slot action space.

Why it matters

The paper gives concrete, reusable recipes for mixing heterogeneous robot datasets: masked unified action slots, flow-based speed alignment, quality metrics for weighting datasets, and a sampling schedule that starts uniform. Its RL stage improves a flow-matching policy without back-propagating RL gradients into the base model, and the same checkpoint family is evaluated on a real humanoid, a real dual-arm cobot and two simulated benchmarks.

Problem

Robot datasets differ in observation layout, action space and sampling rate, and their quality varies from jittery trajectories to blurred frames. Padding all actions into one vector makes a multi-robot policy average incompatible targets, and behaviour cloning alone saturates on long-horizon, contact-rich tasks. Reasoning-heavy VLAs that plan with chain-of-thought add latency that real-time control cannot afford.

Proposed solution

Training runs in stages: a PaliGemma (3B) base, web multimodal grounding on 24M samples (L1), pretraining on about 3,000 hours of multi-robot data (R0), embodiment-specific fine-tuning (R1), and RL alignment (R2) that improves data with a Q-function and learns a noise actor for the flow-matching action expert. A DataQA pipeline filters episodes by tremble, sharpness, diversity and state variance, optical-flow-based resampling equalises motion speed across datasets, and a masked 64-dimensional unified action space with an embodiment prompt replaces zero-padding. At inference the policy predicts episode progress, a Gaussian-mixture OOD detector nudges states back toward the training distribution, and a joint prediction module lifts a VLM-pointed 2D target to 3D and guides flow matching toward it.

Key insights

Zero-padding heterogeneous actions into one vector adds a spurious loss on unused dimensions; masking the loss to a fixed semantic layout removes it and lets one policy keep joint and Cartesian semantics side by side.
Starting the dataset mixture uniform and moving it gradually to target weights stops large datasets from dominating a high-momentum optimiser and washing out rare embodiments.
Mirroring episodes across the robot's bilateral symmetry and time-reversing reversible skills expanded 48 hours of humanoid recordings to 167 effective training hours.
Guiding the flow-matching sampler toward a language-grounded 3D target point raises exact-item picking far more than the policy achieves on its own.
RL alignment that edits trajectories with Q-gradients or learns the source-noise distribution improves success while leaving base-model weights untouched by RL gradients.

Main results

Real AgileX Cobot Magic table cleaning, R0 checkpoint with no fine-tuning against baselines fine-tuned for 20,000 iterations: first-item success 69.5% against 38.4% for AgiBot GO-1 and 35.6% for π0, and mean clearing time 1m35s against 2m59s for π0.
Simpler Google Robot overall average: 60.2 (R0), 61.3 (R1) and 71.8 (R2), against 56.8 for fine-tuned π0, 46.5 for RT-1-X and 33.8 for OpenVLA.
Simpler WidowX average success: 45.0 (R0), 55.2 (R1) and 79.1 (R2), against 73.2 for DB-MemVLA and 27.1 for fine-tuned π0; the text gives the R1-to-R2 gain as an absolute 24%.
CALVIN ABC→D average chain length: 3.6 for π0 R1, 4.18 for Green-VLA R1, 4.53 for Flower R1 and 4.63 for Green-VLA R2.
E-commerce shelf picking: joint prediction guidance raised top-1 success from 62.3% to 95.4% (brand-level), 36.7% to 93.1% (exact SKU) and 10.2% to 72.8% (unseen SKUs).
Green humanoid instruction-following success as plotted in Fig. 12: pick 98%, place in basket 100%, pick from basket 77%, give to user 99%, hand-over 84%, full table cleaning 87%, average 90%.

Limitations

The authors state that performance still depends on retargeting fidelity, residual dataset bias and adequate coverage of dexterous skills.
The trajectory-optimisation RL step is not guaranteed to improve a trajectory, so every optimised trajectory is validated by an operator resetting the scene and re-executing it before it is added to the data.
On CALVIN the R1 model (ACL 4.18) trails Flower R1 (4.53), and the unified action space is not used for this benchmark; R2 reaches 4.63.
Flower, OpenVLA and RT-1-X Simpler numbers are taken from published reports rather than re-run, and the humanoid results are given as percentages without episode counts; the Fig. 12 caption mentions in-domain and out-of-domain settings but the plot shows a single series.
The R0 table-cleaning comparison puts an un-tuned Green-VLA against baselines fine-tuned on task data, and a lightweight reasoning module and multilingual instructions are left to future work.

Field impact

Hugging Face daily papers: 311 upvotes, in the top 5% of robotics papers there over the last six months. The paper links a code repository and a project page.

Abstract

We introduce Green-VLA, a staged Vision–Language–Action framework for real-world deployment on the humanoid Green 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) RL-based policy alignment. Progression builds semantic and physical priors, learns shared affordances, and aligns policies for long-horizon execution beyond behavior cloning. At its core is a unified data and control stack for robot fleets. A scalable data-processing pipeline including DataQA and temporal-alignment filters and synchronizes 3,000 hours of demonstrations; a unified, embodiment-aware action interface enables a single policy to control humanoids, mobile manipulators, and fixed-base arms; and the VLA controller is enhanced with episode-progress prediction, out-of-distribution detection, and a joint-prediction-based guidance module that generalizes to unseen objects. Optimized for the Green humanoid, Green-VLA generalizes in a zero-shot manner to new embodiments and achieves state-of-the-art performance across bimanual systems and benchmarks, with RL alignment providing gains in success rate, robustness, and long-horizon efficiency.

Open Source Robotics