Evaluating 30+ Embodied AI Models Reveals Lack of Robustness

ZeYanjie · x · 2026-07-09

The authors evaluated over 30 frontier embodied AI models, concluding that current general robot policies are still far from achieving robust manipulation in the real world.

To address this, the team built RoboDojo to further research and evaluate the performance of embodied agents in real-world manipulation tasks.

Related event: RoboDojo Benchmarks 30+ Embodied AI Models: General Robotics Still Unstable(2 posts)→

Original post →

More from Embodied

Embodied channel →