Source-linked AI summary
NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks
Chia-Yu Hung, Qi Sun, Pengfei Hong, Amir Zadeh, Chuan Li, U-Xuan Tan, Navonil Majumder, Soujanya Poria
TL;DR
Existing VLA models improve reasoning and task execution but are often large and computationally expensive. NORA uses a 3B-parameter Qwen-2.5-VL-3B backbone with FAST+ action tokenization, and reports improved performance over existing VLA models.
Problem
Existing VLA models often approach or exceed 7B parameters, increasing computational overhead during reasoning and execution.
Method
NORA is a 3B-parameter VLA model built on Qwen-2.5-VL-3B and trained on Open X-Embodiment, using FAST+ to discretize continuous action tokens.
Results
NORA outperforms existing VLA models, with significant task-performance improvements across real-world environments and the LIBERO benchmark.
Takeaways & Limitations
NORA offers an open-source approach that combines strong multimodal reasoning with more efficient action generation for scalable robotic deployment.
Takeaways & Limitations
Action chunking performs worse on WidowX because executing accumulated predicted actions can cause excessively large movements and environmental crashes.
Abstract
from arXiv · showhide
Existing Visual-Language-Action (VLA) models have shown promising performance in zero-shot scenarios, demonstrating impressive task execution and reasoning capabilities. However, a significant challenge arises from the limitations of visual encoding, which can result in failures during tasks such as object grasping. Moreover, these models typically suffer from high computational overhead due to their large sizes, often exceeding 7B parameters. While these models excel in reasoning and task planning, the substantial computational overhead they incur makes them impractical for real-time robotic environments, where speed and efficiency are paramount. To address the limitations of existing VLA models, we propose NORA, a 3B-parameter model designed to reduce computational overhead while maintaining strong task performance. NORA adopts the Qwen-2.5-VL-3B multimodal model as its backbone, leveraging its superior visual-semantic understanding to enhance visual reasoning and action grounding. Additionally, our \model{} is trained on 970k real-world robot demonstrations and equipped with the FAST+ tokenizer for efficient action sequence generation. Experimental results demonstrate that NORA outperforms existing large-scale VLA models, achieving better task performance with significantly reduced computational overhead, making it a more practical solution for real-time robotic autonomy.
1 INTRODUCTION
Existing VLA models improve robotic reasoning but often incur substantial computational overhead because of their large scale. NORA addresses this trade-off with a 3B-parameter design intended to retain task performance while improving efficiency.
- Motivation: Existing VLA models often approach or exceed 7B parameters, increasing computational overhead through intermediate Chain-of-Thought processing.These models combine visual and language understanding to improve task execution accuracy, but processing intermediate reasoning steps raises execution cost.
- NORA: NORA is a 3B-parameter VLA model built on Qwen-2.5-VL-3B and trained on the Open X-Embodiment dataset.The design targets lower computational overhead while maintaining strong task execution capabilities.
- NORA: FAST+ discretizes continuous action tokens to optimize action-sequence generation across robotic tasks.The tokenizer is part of NORA’s efficiency-oriented action-generation design.
- Results: NORA outperforms SpatialVLA in real-world settings without action grids or spatial embeddings.This comparison is reported as part of the model’s introduction and design motivation.
- Results: Experiments cover diverse real-world tasks and the LIBERO simulation benchmark, with significant performance improvements over competitive baselines.The evaluation is intended to assess both practical task execution and benchmark performance.
- Open source: The framework, checkpoints, training strategy, and evaluation protocols are open-sourced to support reproducibility and further research.The release covers the full NORA framework rather than only model weights.
2 PRELIMINARIES
Vision-language models provide strong multimodal understanding and reasoning but are not designed to directly produce embodiment-specific robotic policies. VLA models bridge this gap by mapping visual observations and language instructions to adaptive actions.
- Vision-Language Models: VLM pretraining provides visual semantics that help downstream VLAs ground language in the visual world and translate it into robotic actions.This pretrained knowledge makes VLMs useful backbones for vision-language-action systems.
- Vision-Language Models: VLMs combine image encoders, language-model backbones, and projection networks to reason over text and image modalities.Their pretraining uses diverse multimodal data, including image-text pairs, visual knowledge, grounding, spatial reasoning, and visual question answering.
- Vision-Language-Action Models: VLMs are not inherently designed to generate policies for specific robotic embodiments.VLA models address this limitation by combining visual observations and language instructions to produce adaptive actions across tasks.
- Action Tokenization: Robotic actions commonly represent translation and rotation across multiple degrees of freedom as continuous control signals.Typical dimensions include x, y, z translation and roll, pitch, yaw rotation.
- Action Tokenization: Quantile-based binning maps each action dimension to one of 256 discrete bins for transformer-compatible action prediction.The strategy is intended to limit outlier effects while preserving action granularity.
3 NORA
NORA combines a compact Qwen-2.5-VL backbone, FAST+ action tokenization, and Open X-Embodiment training to generate robotic actions efficiently. Its NORA-LONG variant extends the predicted action horizon to five actions for studying action chunking.
- Architecture: NORA uses Qwen-2.5-VL-3B as its VLM backbone and FAST+ to discretize continuous action tokens.The backbone supplies multimodal reasoning and instruction-following capabilities, while FAST+ supports efficient action encoding and decoding.
- Architecture: NORA autoregressively predicts an action chunk from a natural-language instruction and visual observation frames.The action chunk is encoded as discrete FAST+ tokens conditioned on the concatenated visual and language input.
- Implementation: The implementation uses one visual frame, an action chunk size of 1, 2048 added FAST+ vocabulary tokens, and next-token prediction training.These choices define the basic NORA configuration described in the paper.
- Training: NORA is trained on Open X-Embodiment trajectories spanning different robots and a wide range of tasks.The dataset is used to develop robotic capabilities and generalization across tasks, settings, modalities, and embodiments.
- Training: Training used roughly three weeks on 8 H100 GPUs, totaling approximately 4000 H100 GPU hours.The reported setup used 1.1 million gradient updates, FlashAttention, and bf16 precision.
- NORA-LONG: NORA-LONG predicts an action horizon of 5 actions while sharing NORA’s architecture and pretraining dataset.The variant was pretrained for 900k steps to examine longer-horizon action chunking.
4 EXPERIMENTS
NORA is evaluated as a generalist robotic control model across simulated and real-world tasks, including zero-shot generalization, spatial reasoning, and multi-object manipulation. It generally outperforms baselines, but action chunking and environmental distractions reveal important task-specific limitations.
- Evaluation Setup: NORA is evaluated on zero-shot and fine-tuning-oriented tasks across a WidowX platform and the LIBERO simulation benchmark.The evaluation targets previously unseen tasks and downstream adaptation across diverse environments, embodiments, and task suites.
- Real-World Results: NORA achieves up to 90% success on OOD object-grasping tasks and generally outperforms baselines on spatial-reasoning tasks.SpatialVLA can understand spatial orientation yet fail during grasp execution, while NORA completes the reported object and spatial tasks more reliably.
- Real-World Results: NORA outperforms baselines on multi-object grasping, but success remains below 50% on some multi-object tasks.The narrower advantage and sub-50% success indicate substantial room for improvement when handling multiple objects.
- Simulation Results: 87.9% is NORA-LONG’s highest average LIBERO success rate, while NORA is comparable to OpenVLA on spatial, object, and goal-related suites but weaker on long-horizon tasks.Fine-tuned NORA-LONG achieves the highest overall average across the evaluated methods.
- Robustness: Both OpenVLA and NORA experience significant performance drops when additional objects are introduced as environmental distractions.The result highlights fragility under distractors despite strong performance on the corresponding undisrupted tasks.
- Action Chunking: Action chunking improves LIBERO-Long performance but performs poorly on WidowX, where sequential execution can cause crashes and first-action execution yields 0% on multi-object tasks.With first-action execution, NORA-LONG reaches 80% on carrot placement and 70% on the spatial task, but stops after the first object in multi-object manipulation.
5 RELATED WORKS
Related work develops generalist robot policies, vision-language-action models, spatial and temporal representations, and intermediate reasoning for robotic control. Despite these advances, current VLAs often underuse sequential reasoning capabilities.
- Generalist Robot Policies: Generalist robot policies target diverse tasks across multiple embodiments, extending beyond narrowly defined control settings.
- Vision-Language-Action Models: Vision-Language-Action models integrate robot actions into token spaces to generate low-level controls from multimodal inputs.
- Vision-Language-Action Models: OpenVLA improves accessibility with an open-source 7B-parameter model trained on over 970k real-world robot demonstrations.
- Spatial and Temporal Representations: SpatialVLA encodes 3D spatial information, while TraceVLA represents historical state-action pairs as visual prompts for interactive manipulation.
- Reasoning and Control: CoT-VLA generates future visual goals before action sequences, whereas π0 uses flow matching for general robot control across multiple platforms.
- Open Challenge: Current VLAs often underuse the underlying vision-language model’s ability to reason through sequential steps in complex tasks.
6 CONCLUSION
NORA is a 3B-parameter VLA designed to reduce computational overhead and improve robotic task execution efficiency. It uses Qwen-2.5-VL-3B and FAST+ and is reported to outperform existing VLA models, especially in real-world environments.
- NORA is a 3B-parameter VLA designed to reduce computational overhead and improve execution efficiency.
- NORA uses Qwen-2.5-VL-3B as its backbone and FAST+ to discretize continuous action tokens for faster prediction and efficient action encoding and decoding.
- NORA outperforms existing VLA models, with especially significant task-performance improvements in real-world environments.
A APPENDIX
The appendix contains a placeholder for additional sections.
- The document reserves space for additional appendix sections.
- No specific appendix content is supplied in this passage.
- The appendix placeholder does not report methods, results, or conclusions.