LLARVA: Vision-Action Instruction Tuning Enhances Robot Learning

TL;DR

LLARVA introduces a novel instruction tuning approach with structured prompts and visual trace prediction, significantly improving robot learning across diverse tasks and environments by unifying vision and action representations.

Contribution

The paper presents LLARVA, a new instruction tuning method that leverages structured prompts and visual traces to enhance generalization in robot learning tasks.

Findings

LLARVA outperforms several baselines in robot learning tasks.

The model generalizes well across different environments and robot configurations.

Predicting visual traces improves alignment between vision and action spaces.

Abstract

In recent years, instruction-tuned Large Multimodal Models (LMMs) have been successful at several tasks, including image captioning and visual question answering; yet leveraging these models remains an open question for robotics. Prior LMMs for robotics applications have been extensively trained on language and action data, but their ability to generalize in different settings has often been less than desired. To address this, we introduce LLARVA, a model trained with a novel instruction tuning method that leverages structured prompts to unify a range of robotic learning tasks, scenarios, and environments. Additionally, we show that predicting intermediate 2-D representations, which we refer to as "visual traces", can help further align vision and action spaces for robot learning. We generate 8.5M image-visual trace pairs from the Open X-Embodiment dataset in order to pre-train our model, and we evaluate on 12 different tasks in the RLBench simulator as well as a physical Franka Emika Panda 7-DoF robot. Our experiments yield strong performance, demonstrating that LLARVA - using 2-D and language representations - performs well compared to several contemporary baselines, and can generalize across various robot environments and configurations.