Variants and 10 leaderboards
AI Model Performance tracks what has been published about 15 robot AI models on six criteria, every value checked against its source. Two additions this week.
Variants on every model page. Add-ons, faster settings and fine-tuning fixes built on a model, each result next to the base model in the same paper. The first six papers build on π0, π0.5, SmolVLA, GR00T N1.7 and OpenVLA-OFT. See π0.5’s variants.
10 third-party leaderboards. Groups that run the models themselves, on real robots and in simulation: RoboDojo, RoboChallenge, PhAIL, RoboArena, FlagEval-Robo and five more. Summary numbers only, each one credited and linked to its source.
What their numbers say: on 18 of 22 tests, #1 and #2 are too close to call. The best real-robot success rate is 75%. A production line needs 95% or more.