RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?

За 30 секунд

Что произошло

RoboSPA is a large-scale robotic manipulation dataset and benchmark with 527K trajectories across 280 task variants designed to test Vision-Language-Action models on spatial reasoning and long-horizon planning.

Почему это важно

Robotics engineers building or evaluating VLA models need this to understand whether their systems handle complex scenes and multi-step tasks beyond simple predefined scenarios.

На что обратить внимание

Current VLA models still struggle with complex spatial relations, precise execution, and memory-intensive planning according to experiments, indicating the benchmark reveals significant capability gaps.