PUDA

The hardware harness for Physical AI

PUDA (Physical Unified Device Architecture) connects AI agents to instruments, robots, datastreams, approvals, and audit trails.

puda — zsh

Watch demo Explore demo Read paper

NVIDIA Inception

Member of the NVIDIA Inception program

PUDA is part of NVIDIA Inception, the program that helps AI startups discover their opportunity, build faster, and grow on the NVIDIA platform.

Explore NVIDIA Inception

Make your machines Physical AI ready

PUDA gives teams the runtime layer that makes physical systems AI-ready: execute typed machine actions, preserve the data trail, prove where results came from, analyze outcomes, and close the loop into the next physical action.

Explore demo Read docs

Use cases

The Physical AI lifecycle in one system

Data trails

Capture every action, measurement, and decision

Keep the agent conversation, reasoning, tool calls, machine commands, run state, raw outputs, and derived files aligned in one inspectable and replayable trail.

Closed loop

Turn context into the next action

The AI agent uses this context to run an OODA loop: observe live data, orient with the trail, decide the next action, task, or control decision, and act — with a human in the loop.

post training

Post-train models on real physical runs

Use verified machine actions, sensor measurements, and human-approved outcomes as training signal so post-training is grounded in production physical systems.

Applications

Use cases for Physical AI teams

Robotic workcells

Coordinate arms, grippers, sensors, and inspection tools while preserving the full command and measurement history.

Autonomous test systems

Run test stands, characterization equipment, and validation rigs with repeatable protocols and traceable outputs.

Industrial pilots

Connect AI planning to machines on pilot lines where every action needs operator review, limits, and evidence.

Field operations

Operate distributed physical assets through edge clients while keeping connectivity gaps, approvals, and data capture visible.

Closed-loop optimization

Analyze run results and feed verified outcomes into the next action without losing provenance between iterations.

Human-in-the-loop control

Let agents propose actions while people approve critical steps and audit what happened afterward.

Building Physical AI workflows?

bearspuda@gmail.com