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

TL;DR

Green-VLA presents a staged Vision–Language–Action framework designed to deploy generalist robot policies across multiple embodiments. It couples a five-stage curriculum ($L_0$–$L_1$–$R_0$–$R_1$–$R_2$) with a unified 64-dimensional action space, DataQA-driven data curation, and a flow-matching policy to bridge web-scale priors with real-world robotics data. Key innovations include dynamic embodiment prompts, retargeting to a target humanoid, temporal scale conditioning, and a joint-prediction–driven guidance module, all reinforced by RL alignment to improve long-horizon success and recovery. Empirical results show strong cross-embodiment generalization, large gains from RL refinement on challenging tasks (WidowX, CALVIN), and successful real-robot validation on bimanual and humanoid platforms, underscoring the practicality and scalability of quality-aligned, unified-action robotics foundation models.

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.

Figures (14)

• Figure 1: Green-VLA architecture. A multimodal vision–language model encodes instructions, camera views, and proprioception into tokens that feed a flow-matching action expert. A high-level task planner decomposes user goals into subtasks, queries the VLA loop, and uses auxiliary signals (episode end, OOD, and JPM-based guidance for precise target points) to ensure safe, instruction-faithful execution across embodiments.
• Figure 2: Green-VLA’s robot-specific training stages use visual question answering (VQA) and robotics data and enable robot adaptation and specialization for new embodiments, spatial reasoning, task generalization, dexterous manipulation, and failure recovery.
• Figure 3: Datasets mixture used in L1 training phase. Left: distribution of sample counts across sub-datasets. Right: sampling weight allocation across categories. The data corpus integrates diverse web sources covering spatial reasoning, pointing, robotics-related VQA, and multi-view QA.
• Figure 4: Left: Dataset sampling rates used during the R0 phase of GreenVLA training. Right: Number of data samples (frames) per dataset, illustrating relative temporal coverage. The corpus combines large-scale open datasets (e.g., AgibotWorld, DROID, Galaxea) with internally collected humanoid and dexterous-hand data.
• Figure 5: Overview of the data pipeline for robot learning. Data collection and processing loop integrating robot-side teleoperation, cloud-based data verification, open-source dataset mining, and model training. The pipeline supports iterative model updates via RL fine-tuning and feedback from real-robot deployments.