Towards Fast, Memory-based and Data-Efficient Vision-Language Policy

TL;DR

LiteVLP tackles the challenges of expensive inference, domain shift, and memory limitations in vision-language policies for robotics by leveraging a lightweight 1B-parameter VLM backbone, memory-aware visuomotor instruction tuning, and a novel multi-observation compression module. The approach enables memory of past and future experiences while significantly reducing input token counts, boosting both training efficiency and inference speed. On VIMA-Bench, LiteVLP achieves competitive or superior task success using only a fraction of the data and exhibits strong robustness to localization noise, with notable gains in long-horizon manipulation. The work demonstrates that compact multi-modal models can deliver real-time, data-efficient robotic reasoning, paving the way for scalable deployment in real-world settings.

Abstract

Vision Language Models (VLMs) pretrained on Internet-scale vision-language data have demonstrated the potential to transfer their knowledge to robotic learning. However, the existing paradigm encounters three critical challenges: (1) expensive inference cost resulting from large-scale model parameters, (2) frequent domain shifts caused by mismatched data modalities, and (3) limited capacity to handle past or future experiences. In this work, we propose LiteVLP, a lightweight, memory-based, and general-purpose vision-language policy generation model. LiteVLP is built upon a pre-trained 1B-parameter VLM and fine-tuned on a tiny-scale and conversation-style robotic dataset. Through extensive experiments, we demonstrate that LiteVLP outperforms state-of-the-art vision-language policy on VIMA-Bench, with minimal training time. Furthermore, LiteVLP exhibits superior inference speed while maintaining exceptional high accuracy. In long-horizon manipulation tasks, LiteVLP also shows remarkable memory ability, outperforming the best-performing baseline model by 18.8%. These results highlight LiteVLP as a promising model to integrating the intelligence of VLMs into robotic learning.

Figures (7)

• Figure 1: Comparative performance of vision-language policies. The x-axis represents the average success rate on VIMA-Bench, and the y-axis represents the inference speed evaluated on same devices. The bubble diameter indicates the number of model parameters
• Figure 2: Overall framework of LiteVLP. The LiteVLP initiates with multi-observation compression and then projects the image features into the same dimensional space as the text features. Subsequently, the image tokens are interleaved with text tokens and processed by a large language model to generate a text output that includes the end-effector's action. Of note, during the fine-tuning stage, the parameters of the ViT are frozen, while the length embedding, the MLP projector and the large language model are trained.
• Figure 3: Simple visualization of MOC's effect. The light gray image patches indicate unchanged areas between consecutive images, which will be reduced in the sequence of image patches.
• Figure 4: Example of D-inBC dataset format. The D-inBC dataset includes the task description and the description of reference images. A data converter processes instructions, while an object detector extracts the locations of each object to form the image descriptions.
• Figure 5: Success rates of all tasks. Note that the difficulty level L3 doesn't include sweep without exceeding task.