RoboChallenge Annual Report on Large-Scale Real Robot Benchmark

In early 2026, RoboChallenge released the world’s first large-scale evaluation leaderboard for Vision-Language-Action (VLA) models based on real robot environments. The data shows exponential growth in both registered users and evaluation submissions, highlighting that real robot large-scale evaluation is becoming a critical requirement for embodied intelligence development.

Currently, the top-ranked model achieves around 62% success rate on the Table30 task suite, while performance on fine-grained manipulation tasks remains below 15%. This indicates that embodied foundation models are still at a pivotal stage, transitioning from semantic understanding toward breakthroughs in complex physical interaction.

The platform has attracted developers from multiple countries, including China, the United States, Singapore, and Japan, forming an increasingly global ecosystem. Notably, the majority of top-performing models on the leaderboard are open source, underscoring that open collaboration is becoming a key driver in advancing embodied intelligence.