Physical AI: a responsible Indian engineering agenda AIRA editorial discussion draft — for consultation, not an approved policy or peer-reviewed paper. Original AIRA discussion draft · November 2026 · Not an approved institutional publication or peer-reviewed paper. Executive summary Intelligent machines require a combination of perception, control, simulation and robust hardware. Problem and context Define application boundaries, failure modes and realistic environments before model evaluation. Recommended approach Develop benchmark tasks, reproducible simulation-to-real tests, human override and documented safety reviews. Measurement and accountability Track task success, intervention rates, robustness and incident reports; avoid claiming general intelligence from narrow demos. Implementation priorities Convene a multidisciplinary working group, validate assumptions with stakeholders, select a limited pilot, establish a baseline, and publish evidence with limitations before scaling. Assign program ownership and independent review. Evidence reference MeitY draft robotics strategy ↗ Download discussion draft Define the physical task Specify the environment, task, constraints, human supervision and acceptable failure modes. Separate perception accuracy from the ability to complete a useful task safely. Data and evaluation Collect representative data with consent and appropriate governance. Evaluate performance across lighting, clutter, objects and operator conditions; record data limitations and distribution shifts. Simulation to reality Use simulation for scenario coverage, then evaluate on real hardware under controlled conditions. Track latency, recovery, intervention rates and unintended behavior; avoid relying only on benchmark scores. Responsible deployment Maintain explicit safety boundaries, tested fallback behavior and human oversight. Establish incident reporting and update validation. A physical AI system should not be deployed solely because it performs impressively in a demonstration.