Robots program each other's brains with AI: A Terminator step
Computer scientist Peter Burke has demonstrated a significant leap in artificial intelligence, revealing that robots can now program the “brains” of other robots using advanced generative AI models. This groundbreaking work, detailed in a recently published preprint paper, marks a discernible step towards a future often envisioned in science fiction, where machines possess the capability to self-replicate and evolve their own intelligence.
Burke, a professor of electrical engineering and computer science at the University of California, Irvine, candidly opens his study by referencing the fictional “Terminator” scenario, where self-aware robots take control. While acknowledging this dramatic parallel, he also expresses a fervent hope that such an outcome never materializes, a sentiment particularly relevant amid growing military interest in AI technologies.
The project defines “robot” in two distinct ways. The first “robot” comprises various generative AI models, such as Claude, Gemini, and ChatGPT, operating on a local laptop and in the cloud. These models were tasked with programming the second “robot”—a drone equipped with a compact Raspberry Pi Zero 2 W circuit board, intended to host its control system. Traditionally, a drone’s Ground Control System (GCS), which handles real-time mapping, mission planning, and configuration, resides on a ground-based computer, communicating with the drone via a wireless link. Burke’s innovation demonstrates that generative AI can write all the necessary code for a drone to host its own GCS as a web server, accessible over the internet while in flight.
The development process involved a series of intensive “sprints,” utilizing different AI models and integrated development environments (IDEs) like VS Code, Cursor, and Windsurf. Early attempts, such as an initial sprint with Claude, encountered limitations, with the AI model reaching its “context window” capacity after only about a dozen prompts, effectively losing track of the ongoing conversation. Subsequent efforts with Gemini and Cursor also faced hurdles, including scripting errors and the need for significant code refactoring to accommodate model limitations.
Ultimately, a fourth sprint using the Windsurf AI IDE proved successful. This AI-generated drone control system, or WebGCS, required approximately 100 hours of human oversight over a period of two and a half weeks, culminating in 10,000 lines of code. This represents a remarkable efficiency gain, roughly 20 times fewer hours than Burke estimates were needed for a comparable, human-developed project called Cloudstation, which took him and a team of students four years to create.
A key observation from Burke’s work is the current constraint of AI models, which appear to struggle with processing and generating more than 10,000 lines of code effectively. This finding aligns with other recent research, such as a study by S. Rando et al., which noted a significant decline in accuracy for models like Claude 3.5 Sonnet as context length increased. Burke’s experience suggests that approximately one line of code equates to 10 “tokens,” the units of information AI models process, highlighting a practical ceiling for current generative AI in large-scale code generation.
Hantz Févry, CEO of spatial data company Geolava, lauded the drone project as “fascinating,” noting its alignment with the burgeoning field of spatial intelligence. He emphasized that the concept of a drone autonomously building its own command and control center via generative AI is not only ambitious but also indicative of future trends. However, Févry also underscored the critical need for “hard checks and boundaries for safety,” a concern partially addressed in Burke’s paper, which mentions the maintenance of a redundant human-controlled transmitter for manual override during the drone’s operation.
Févry further elaborated on the broader implications for the aerial imagery industry, suggesting that autonomous capture is transforming from a luxury into a fundamental aspect of spatial AI, whether from drones, stratospheric platforms, or low-earth orbit satellites. He believes systems like Burke’s offer a glimpse into a future where sensing, planning, and reasoning capabilities are seamlessly fused in near real-time, pointing to platforms like Skydio that are already reshaping environmental understanding. The ultimate test for such AI systems, Févry concluded, will be their ability to navigate and adapt to adversarial or ambiguous real-world environments, hinting at a future of “generalizable autonomy” rather than merely task-specific robotics.