Anthropic sets rules for agents that drive microscopes and robot arms
Anthropic's Model Hardware Standard defines how agents may control microscopes, laboratory equipment, factory machinery and robot arms, while acknowledging risks to people and equipment. The control problem is becoming tangible: Hugging Face opened pre-orders at $399 for a 25-centimetre open-source robot whose simulated behaviours can move onto hardware, while Nvidia published a workflow that adapts one navigation policy to different robot bodies.
Artificial Intelligence··Night
Physical limits for a software agent
Anthropic introduced the Model Hardware Standard, rules for how agents interact with microscopes, liquid-handling equipment, quantum-computing hardware, manufacturing machines and robot arms. Developed with manufacturers, the company positions it as a physical counterpart to the Model Context Protocol that connects agents to software tools. Anthropic acknowledges risks including damaged equipment, injured people and biological-weapons development, and says guardrails in its own models should prevent malicious use.[1]
From simulation to a 25-centimetre robot
Hugging Face opened pre-orders at $399 for Microduck, a 25-centimetre open-source robot with a camera, lidar sensors and inertial measurement units. Its development kit and reinforcement-learning stack are published on GitHub, and behaviours trained in simulation can be deployed directly to the hardware. Microduck can waddle, crouch, recover from a fall and lift objects weighing up to 800 grams. The company expects shipping to begin before Christmas.[2]
One policy, different robot bodies
Nvidia's open-source COMPASS workflow adapts the pretrained X-Mobility navigation policy with layers specific to a new robot body. Reinforcement learning tests the robot on measures such as reaching a goal, falling and path length; agents handle steps from environment validation to runtime packaging, with human approval gates between them. The reference implementation trains Boston Dynamics' four-legged Spot across warehouse, generated indoor and reconstructed real-world scenes.[3]
