ControlVLA: Few-shot Object-centric Adaptation for Pre-trained Vision-Language-Action Models

TL;DR

ControlVLA introduces a novel few-shot learning framework that adapts pre-trained vision-language-action models to object-centric robotic manipulation tasks with minimal demonstrations, outperforming traditional methods.

Contribution

It proposes a zero-initialized projection layer approach for efficient fine-tuning of pre-trained models to object-centric tasks in robotics.

Findings

Achieves 76.7% success rate with only 10-20 demonstrations.

Outperforms traditional methods requiring over 100 demonstrations.

Extensible to long-horizon tasks and robust to unseen objects.

Abstract

Learning real-world robotic manipulation is challenging, particularly when limited demonstrations are available. Existing methods for few-shot manipulation often rely on simulation-augmented data or pre-built modules like grasping and pose estimation, which struggle with sim-to-real gaps and lack extensibility. While large-scale imitation pre-training shows promise, adapting these general-purpose policies to specific tasks in data-scarce settings remains unexplored. To achieve this, we propose ControlVLA, a novel framework that bridges pre-trained VLA models with object-centric representations via a ControlNet-style architecture for efficient fine-tuning. Specifically, to introduce object-centric conditions without overwriting prior knowledge, ControlVLA zero-initializes a set of projection layers, allowing them to gradually adapt the pre-trained manipulation policies. In real-world experiments across 6 diverse tasks, including pouring cubes and folding clothes, our method achieves a 76.7% success rate while requiring only 10-20 demonstrations -- a significant improvement over traditional approaches that require more than 100 demonstrations to achieve comparable success. Additional experiments highlight ControlVLA's extensibility to long-horizon tasks and robustness to unseen objects and backgrounds.