InternRobotics/InternVLA-M1

What it solves

InternVLA-M1 addresses the challenge of creating generalist robot policies that can understand complex visual scenes and execute precise actions. It aims to bridge the gap between high-level vision-language understanding and low-level robotic control by providing a framework that can be easily extended and trained on diverse datasets.

How it works

The framework utilizes a spatially guided vision-language-action (VLA) approach. It integrates a language head and an action head within a unified system, allowing for dual supervision during training. By leveraging large-scale multimodal pretraining (such as Qwen2.5-VL) and applying spatial prompt fine-tuning, the model transfers visual and spatial priors to robotic tasks, enabling fast convergence without requiring separate action pretraining.

Who it’s for

• Developers wanting to use open-source Vision-Language Models (VLMs) for robot control.
• Teams co-training action datasets alongside multimodal vision-language data.
• Researchers studying VLA architectures and training strategies.

Highlights

• Dual-System Architecture: Combines language and action heads for collaborative training.
• Modular Design: Decouples model architecture, training data, and evaluation pipelines for easier debugging and extension.
• Hardware Flexibility: Supports NVIDIA CUDA, Intel GPUs (XPU), and CPUs.
• High Performance: Demonstrates SOTA-level convergence and strong results across benchmarks like WindowX, Google Robot, and LIBERO.