BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation

2025-06-09

AI Summary

To address the challenge of deploying large-scale Vision-Language-Action (VLA) models on resource-constrained robotic platforms, this paper proposes the first fully parameterized 1-bit VLA model with {-1, 0, 1} ternary weights. Methodologically, it introduces three key innovations: (i) the first full-parameter 1-bit ternarization of VLA models; (ii) a distillation-aware training strategy that compresses the visual encoder to 1.58 bits while preserving cross-modal representation alignment; and (iii) integration of a lightweight visual encoder, knowledge distillation, and robot manipulation task fine-tuning. Evaluated on the LIBERO benchmark, the proposed model achieves performance comparable to the 4-bit quantized OpenVLA-OFT, yet reduces memory footprint to just 29.8% of the latter’s—substantially enhancing feasibility for edge deployment.

Technology Category

• Computer Vision: Large Vision Models
• Intelligent Robots: Learning & Optimization for ROB
• Machine Learning: Learning on the Edge & Model Compression

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Vision-Language-Action (VLA) models have shown impressive capabilities across a wide range of robotics manipulation tasks. However, their growing model size poses significant challenges for deployment on resource-constrained robotic systems. While 1-bit pretraining has proven effective for enhancing the inference efficiency of large language models with minimal performance loss, its application to VLA models remains underexplored. In this work, we present BitVLA, the first 1-bit VLA model for robotics manipulation, in which every parameter is ternary, i.e., {-1, 0, 1}. To further reduce the memory footprint of the vision encoder, we propose the distillation-aware training strategy that compresses the full-precision encoder to 1.58-bit weights. During this process, a full-precision encoder serves as a teacher model to better align latent representations. Despite the lack of large-scale robotics pretraining, BitVLA achieves performance comparable to the state-of-the-art model OpenVLA-OFT with 4-bit post-training quantization on the LIBERO benchmark, while consuming only 29.8% of the memory. These results highlight BitVLA's promise for deployment on memory-constrained edge devices. We release the code and model weights in https://github.com/ustcwhy/BitVLA.

Problem

• Reducing VLA model size for resource-constrained robotics
• Applying 1-bit pretraining to Vision-Language-Action models
• Compressing vision encoder memory footprint efficiently

Innovation

• 1-bit ternary VLA model for robotics
• Distillation-aware training for vision encoder
• Memory-efficient deployment on edge devices