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

TL;DR

BitVLA introduces a 1-bit vision-language-action model for robotics manipulation to tackle memory and compute constraints on edge devices. It pairs a ternary-parameter 1-bit LLM with a distillation-aware, 1.58-bit vision encoder, achieving competitive performance on the LIBERO benchmark with a fraction of the memory required by 4-bit PTQ baselines. The approach demonstrates that aggressive quantization, coupled with latent-representation alignment, can preserve downstream capabilities for manipulation and visual reasoning while enabling deployment on resource-constrained hardware, and it extends to zero-shot VQA with minimal accuracy loss. The work includes extensive ablations and qualitative analyses and releases code and weights to support edge-friendly VLA deployment in robotics.

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.

Figures (6)

• Figure 1: Comparison between BitVLA and OpenVLA-OFT with 4-bit post-training quantization in terms of end task performance and memory footprint. We report the average success rate on LIBERO benchmark.
• Figure 2: The overview of the training of BitVLA. We first train a vision-language model using a 1-bit LLM bitnet2b combined with a full-precision vision encoder. Then we apply distillation-aware training to quantize the vision encoder's weights to 1.58-bit precision. Finally, BitVLA is adapted to specific robotics tasks through OFT fine-tuning openvla-oft.
• Figure 3: The overview of the distillation-aware training. The original full-precision encoder serves as the teacher model to ensure better alignment of the latent representations.
• Figure 4: The overview of LIBERO benchmark task suites. It has four different dimensions to evaluate the generalization and performance of robotics manipulation models.
• Figure 5: Left: distribution of failure types across each task suite on the LIBERO simulation environment. Right: typical examples for (a) spatial localization discrepancy, (b) goal misunderstanding and (c) trajectory planning failure.