AI can write your code. It still can't run your project.
A new paper surveyed teams competing in the European Rover Challenge (ERC), an annual contest where student groups design and build rovers under tight deadlines and heavy interdependence between subsystems. Researchers sent a 40 question, role adaptive survey to ERC 2025 participants and got 104 responses from 14 teams, covering team structure, task management, communication, and integration practices. The results point to familiar failure modes: thin documentation, fuzzy requirements, communication that splinters across tools, task tracking done informally, and significant rework when subsystems finally get integrated. From those findings, the authors sketch an assistant architecture meant to handle requirement tracking, communication summarization, integration risk flagging, and ongoing knowledge capture, tying together user interfaces, credential management, and a menu of specialized AI services.
Most AI tooling sold to engineering teams targets individual tasks, autocomplete for code, drafting docs, searching a knowledge base. This paper's finding is that the expensive failures happen at the process level, in the gaps between people and subsystems, which is exactly where today's AI tools don't reach. That's a useful corrective for anyone assuming a coding assistant plus a chatbot equals a coordinated team.
The architecture here is still a proposal, not a working system, so treat this as a requirements document with good instincts rather than a product review.