RealMirror: An Open-Source Vision-Language-Action Platform for Embodied Intelligence

Tai et al. propose RealMirror, an open-source vision-language-action (VLA) platform for embodied intelligence that integrates efficient data collection, standardized evaluation, and environment reconstruction based on 3D Gaussian Splatting. The platform allows researchers to develop and iterate VLA models without physical robot hardware, and provides a VLA benchmark containing multiple scenarios and standardized trajectories. The paper demonstrates zero-shot sim-to-real transfer, where models trained purely in simulation can complete real robot tasks without fine-tuning. From a safety perspective, Failure-First notes that adversarial attacks validated in simulation (such as the backdoor attack documented by DropVLA) may also transfer directly to physical robots without an adaptation phase, and asks whether this benchmark can be extended to safety evaluation under adversarial conditions.