Spirit AI has open-sourced Spirit v1.5, the top-ranked model on the RoboChallenge Table30 benchmark.
The AMW Read
The release of a top-ranked open-source VLA model updates the robotics landscape by validating a shift toward unscripted, goal-driven data scaling for embodied intelligence.
NoveltySignificance
Robotics · Player MapScaling Laws
Spirit AI has open-sourced Spirit v1.5, the top-ranked model on the RoboChallenge Table30 benchmark. By utilizing a unified Vision-Language-Action architecture, the model shifts the training paradigm from scripted demonstrations to unscripted, goal-driven data. Results show that task diversity, not task purity, significantly improves transfer efficiency and mastery of novel tasks during fine-tuning. This release of weights and code marks a systemic shift toward scalable, general-purpose embodied intelligence. 🤖🚀
