Spirit AI's v1.5 robot model tops the RoboChallenge real-robot benchmark and is released open source

Spirit AI (Qianxun Intelligence)/Spirit VLA models

Spirit AI said its Spirit v1.5 model ranked first on RoboChallenge, a benchmark that scores robot models on 30 everyday tasks run on physical robots. It scored 66.09 with a 50.33% task success rate, ahead of Physical Intelligence's pi0.5, and was the only model above 50%. The company open-sourced the model and related resources.

• Total score 66.09; task success 50.33%.
• Ahead of Physical Intelligence's pi0.5 on the same benchmark.
• RoboChallenge covers 30 real-robot tasks such as placing objects and using tools.
• The model and related resources were released openly.