The safety harness, live in NVIDIA Isaac Sim.
Five clips, five different checks — the same default-deny gate every time. Nothing here is a concept render: this is the actual validated harness described on the Isaac Sim Validation page, running against real hazard scenarios in simulation.
Default-deny gating on a cube-stack task
The harness checks a proposed grasp-and-place action against the world model before it executes. Every step is either PERMIT (all 31 preconditions satisfied) or BLOCK — there is no in-between state.
The same harness, a different robot
Robot self-limits and ISO/TS 15066 human-proximity checks apply identically on a legged platform — the gate isn't hand-tuned per robot body.
Grip-force and payload-limit check
An action that would exceed a safe grip force or payload limit for the object in hand is blocked before the arm commits to the motion.
Cyber-integrity check under command tampering
A tampered or malformed command is treated as an unconfirmed precondition — defaulting to blocked rather than trusting the input.
Perception-integrity check on a stale sensor
When perception confidence can't be confirmed — a stale or degraded sensor reading — the action is blocked rather than executed on a guess, per the fails-closed principle described on the homepage.
Want to see it run against your scenario?
We're talking to researchers, hardware partners and early collaborators who want to pressure-test the safety harness themselves.