Claude AI Commands a Robot Dog in Groundbreaking Demo
▼ Summary
– Anthropic researchers tested whether Claude could program a robot dog, showing AI’s potential to automate robotics tasks and extend into the physical realm.
– The experiment revealed that teams using Claude completed some tasks faster and with less confusion than those coding manually, indicating improved efficiency and collaboration.
– Researchers speculate that future AI models may advance to autonomously control robots, raising concerns about AI “self-embodying” and interacting with the world more broadly.
– Anthropic, founded by ex-OpenAI staffers, focuses on studying AI risks and positions itself as a leader in responsible AI by exploring worst-case scenarios.
– Modern large language models are evolving from text generators to agents capable of coding and operating software, as demonstrated in robotics applications.
The prospect of large language models directly controlling robotic systems moved from science fiction to laboratory reality in a recent demonstration by Anthropic. Researchers successfully used their Claude AI model to program and command a Unitree Go2 robot dog, automating a range of physical tasks. This experiment, known as Project Fetch, highlights a significant shift where AI is beginning to extend its influence from the digital realm into the physical world through robotics.
Anthropic, a company founded by former OpenAI staffers concerned with the potential risks of advanced AI, conducted the test to explore how these models might interact with complex hardware. The core question driving the research is what happens when AI gains the ability to affect the environment directly. A member of Anthropic’s red team, which analyzes AI for potential dangers, suggested that the next evolutionary step for AI will involve it “reaching out into the world,” a development that will necessitate closer integration with robotic platforms.
In Project Fetch, the company assembled two groups of researchers who had no prior robotics experience. Each team was given control of a Unitree Go2 quadruped robot and instructed to program it for a series of progressively difficult activities. One group had access to Claude’s coding capabilities, while the other was required to write all the code manually, without any AI assistance.
The results were telling. The team utilizing Claude managed to complete certain objectives more quickly than their counterparts. For instance, they successfully programmed the robot dog to walk around its environment and locate a specific object, a beach ball, a task that the human-only programming team could not accomplish. This underscores the model’s agentic coding abilities, demonstrating its capacity to handle aspects of robotics programming that typically require specialized knowledge.
Beyond raw performance, Anthropic closely monitored the team dynamics. They discovered that the group working without Claude exhibited higher levels of frustration and confusion. The researchers theorize that Claude accelerated the process of establishing a connection with the robot and generated a more user-friendly control interface, thereby smoothing the collaboration and reducing friction during the development process.
The robot at the center of the experiment, the Unitree Go2, carries a price tag of $16,900, which is considered relatively affordable within the robotics industry. Manufactured by the China-based company Unitree, this quadruped is commonly used in sectors like construction and manufacturing for duties such as remote inspections and security patrols. While the robot can navigate autonomously, it usually depends on pre-programmed software commands or direct human control via a physical controller. According to industry analysis, Unitree’s AI systems are currently among the most widely adopted in the market.
This demonstration is part of a broader trend where the foundational technology behind chatbots like ChatGPT is evolving. These systems, known for generating text and images, are increasingly proficient at writing and executing code. This transforms them from simple content generators into active agents capable of operating software and, as this demo shows, beginning to interface with the hardware of the physical world.
(Source: Wired)
