MemoryVLA is a novel open-source Vision-Language-Action (VLA) model that integrates a memory module to allow robots to recall and learn from previous failures. This post provides a detailed overview of the model's architecture, including its perception-cognition memory system, and explains how it was trained and can be deployed locally. The key innovation is the ability to store and retrieve experiences, enabling robots to avoid repeating mistakes and improve over time. This represents a significant advancement in making robots more adaptive and reliable in real-world environments. For AI and robotics researchers, MemoryVLA offers a practical framework for building more intelligent autonomous systems. The post also covers the implications for ICLR 2026 and the broader field of embodied AI.
An open-source VLA model with memory that enables robots to learn from past failures, advancing autonomous systems.