Anthropic proposes plumbing spec to link AI agents to lab kit and robots

Anthropic has introduced a protocol for safely operating physical devices with AI agents, despite the inherent uncertainty in predicting model behavior. Named the Model Hardware Standard (MHS), this protocol is akin to the Model Context Protocol (MCP), which enables AI models to connect to data sources. Anthropic envisions MHS allowing Claude and other models to control laboratory equipment, factory machinery, and robots.

The protocol is presently being tested at the Howard Hughes Medical Institute's Janelia Research Campus in Maryland and can be applied to other labs and industrial facilities.

Setting up hardware integration typically takes weeks or months, with devices often lacking communication between them, requiring specialists to create custom integrations. MHS aims to be a universal translation layer, reducing integration time to hours or minutes. It utilizes a programmable interface, with many industrial machines already offering interfaces to expose controls and data, though the diversity of such devices makes it challenging for developers to create connections. MHS acts as a universal translation layer for these diverse devices.

The driver software employs limited primitives like read and write, similar to how simple tools like Bash can power AI agents. The driver makes devices discoverable in a standard format, supports tags conveying device functions, and allows users to provide data by conversing with the model during setup. These tags enable the driver to generate a reference file detailing device characteristics.

AI agents can interact with devices using three control paths: the MCP, command line interface, and API code, enabling them to carry out commands, monitor test results, or adjust knobs and dials.

Biotechnology company Genentech has utilized MHS for drug-discovery experiments with real-time error handling, while quantum computing firm QuEra improved laser stabilization for its machines by 99.3 percent. Companies such as AWS, Automata, Danaher, Doosan Robotics, MBF Bioscience, Qiagen, Tecan, and Universal Robotics are also planning to support MHS.

However, further safety evaluations and protections are necessary, as LLMs still lack physical intuition, having learned about the physical world primarily from text and images.