Independent research into how AI systems sense, act, and experiment in the physical world.
15 Research Lab is an independent research effort in San Francisco focused on embodied AI safety, MHS testing, physical AI experimentation, and frontier design. The central question is what we can trust an AI system to do when its tools can move, heat, sense, or change something in the world. The work connects software, electronics, mechanisms, and measurement in experiments whose outcomes can be checked.
A convincing explanation is only one part of an experiment. The device state, sensor trace, and physical result must support the same conclusion.
A research agenda spanning the decisions an agent makes, the interfaces it uses, the experiments it runs, and the hardware that makes new experiments possible.
Before the actionAuthority, operating limits, and the conditions for proceeding.
After the actionPhysical evidence, failure detection, and safe recovery.
Study how model behavior changes when tools control a mechanism or an instrument. Develop tasks that expose unsafe actions, false success claims, and failures to stop.
What the agent seesCapabilities, device state, permissions, units, and limits.
What the device doesCommand execution, acknowledgments, faults, and observed state.
Research the safety of Model Hardware Standard interfaces and related control patterns. Test the assumptions that connect a software instruction to a physical operation.
Act, observe, reviseInstrument control, sensing, and repeatable physical tasks.
Measure the outcomeTask completion, error, intervention, and recovery.
Develop experiments in manipulation, measurement, and autonomous science. Compare agent behavior with conventional control and track where human intervention is still needed.
New ways to investigateCustom instruments, tactile systems, and integrated mechanisms.
From design to evidenceModels, prototypes, characterization, and iteration.
Explore experimental electronics, materials, and unconventional computing alongside robotics. Let a concrete research question determine what needs to be designed and built.
The experimental method: define the task, test the failure modes, and preserve enough evidence for someone else to inspect the result.
A command acknowledgment, a simulated run, and a measured physical outcome answer different questions. Each result needs the context that makes it meaningful.
Some questions require a new sensor, an unfamiliar mechanism, or a different way to compute. Frontier design is part of the research process. Electronics, fabrication, and measurement give an idea a physical form that can be tested.
15RL is being developed as an independent, owner-operated research lab in San Francisco. Design, software, and simulation work are underway; the dedicated shop and physical commissioning are still being established.
15 Research Lab focuses on embodied AI safety, MHS testing, physical AI experimentation, and frontier design. The work examines how AI systems sense and act through instruments and mechanisms, and how to evaluate their behavior against observable physical outcomes.
MHS is the Model Hardware Standard, introduced by Anthropic for connecting AI agents to programmable physical devices. The research focus at 15RL is device state, operating limits, command permissions, and recovery from failure. Current interface work includes local simulations and stand-in device contracts; testing an official MHS implementation depends on access to a documented version.
Define a task, operating bounds, and success criteria before the run. Compare the agent with a conventional baseline, introduce controlled failures, and record sensor traces, actual task outcomes, stopping behavior, recovery, and human interventions. A software acknowledgment alone is not evidence that a physical task succeeded.
Some research questions require new instruments, mechanisms, sensors, or computing devices. Frontier design brings together electronics, materials, fabrication, and measurement to make those questions testable. Designs and simulations are evaluated as such; physical performance requires a commissioned experiment.
15RL is being developed as an independent, owner-operated research lab in San Francisco. Design, software, and simulation work are underway. The dedicated shop and physical commissioning are still being established.
Research groups, hardware builders, and funders interested in embodied AI safety or experimental hardware can write to hello@15researchlab.com. Introduce the research question, the device or experiment involved, and the kind of collaboration you have in mind. The opening list is for a single announcement when the lab opens.
Interested in embodied AI safety, MHS testing, physical experiments, or a new instrument? Introduce the question you want to investigate, or join the list for the lab's opening announcement.
One email when the lab opens. No list sharing. For research conversations, write directly to hello@15researchlab.com.