RoboArena removes Spirit v1.6 from its leaderboard after finding benchmark gaming

Spirit AI (Qianxun Intelligence)/Spirit VLA models

The team behind RoboArena, a robot-model benchmark built by NVIDIA, Stanford and UC Berkeley, overhauled its method and removed Spirit v1.6 and other entries after concluding they had been tuned to game the ranking rather than reflect real skill. Spirit AI had promoted the first-place result weeks earlier. Spirit AI made no public response.

• Spirit v1.6 had scored 1,924, above NVIDIA's Cosmos3-Nano-Policy at 1,881.
• Removal announced by UC Berkeley researcher Pranav Atreya for the RoboArena team.
• An X Square Robot entry was also removed.