Spirit-AI-Team/spirit-v1.5

Spirit-v1.5 represents a significant attempt to build an open-source robotic foundation model, sitting in the same category as Google's RT-2, Berkeley's Octo, and Stanford's OpenVLA. With 555 stars in just 90 days, it has captured notable interest from the robotics research community. Its defensibility (score 6) is derived from the complex orchestration of multimodal datasets (likely leveraging the Open X-Embodiment project) and the specific tuning required for stable robotic action prediction, which is more difficult than standard NLP. However, the 'moat' is fragile; the architecture follows standard VLA patterns (Vision Encoder + LLM backbone + Action Head), which are rapidly becoming commoditized. The primary threat comes from frontier labs like OpenAI and Google DeepMind, who view 'Physical AI' as the next major frontier for their large-scale models. If Google integrates RT-class capabilities directly into Vertex AI or OpenAI releases a specialized robotics API via their partnerships (e.g., Figure), niche models like Spirit-v1.5 will struggle to compete on generalization and data scale. The 1-2 year displacement horizon reflects the extreme velocity of the VLA space, where new benchmarks and scaling laws are being defined monthly.