We stand on the work of researchers and platform builders. Nothing here is a rivalry. We are building one specific, testable layer: an immutable hardware core with its own sensing, a fail-safe ladder and sector profiles, designed to work alongside the platforms the field already uses.
| Work | What it shows | How Inztinct builds on it |
|---|---|---|
| RoboPAIR arXiv 2410.13691 | Jailbreak attacks can make LLM-controlled robots take harmful physical actions. | We take the warning seriously. We assume the planner can be compromised and keep the last line of defence outside it. |
| Black-Box Simplex arXiv 2102.12981 | A verified controller can take over from an unverified one at run time. | We build on this pattern for physical AI, adding fixed hardware, independent sensing and sector profiles. |
| Zhang and Lu arXiv 2307.10690 | An earlier use of an "instinct layer" for intelligent agents. | We credit earlier uses of the term. In our work instinct means a fixed hardware boundary, not a learned module. |
| Hrstek Bridging Intelligence and Instinct | A framing of fast instinct beside slower intelligence for autonomous systems. | We build on this framing with an immutable core, a fail-safe ladder and a certification path. |
| NVIDIA Halos | A full-stack safety platform for robots and vehicles, with isolation, monitoring and certified processes. | Complementary. Halos supplies certified compute and processes. ZIV is designed to run on Jetson-class compute, with the safety core beneath as an AI-independent fallback. |
| Dual-system robot models Figure Helix, NVIDIA GR00T N1 | A fast reflex model beside a slower reasoning model. | We credit this work for the System 1 and System 2 vocabulary. We call the learned fast layer intuition and keep instinct for a fixed core. |
We thank the authors of RoboPAIR, Black-Box Simplex and the instinct-layer work cited above, the teams behind Figure Helix and NVIDIA GR00T N1 for the System 1 and System 2 vocabulary, and NVIDIA for its public safety work on Halos and Jetson. The TU Berlin researchers who published the voltage-glitch attack on a car root of trust taught us as much as any design paper. NVIDIA, Jetson and Halos are trademarks of NVIDIA Corporation. Figure and Helix are trademarks of their owners. Inztinct is not affiliated with or endorsed by any company named on this page. Descriptions are our reading of public sources and may be corrected on request.
Nothing below is published yet. We will release preprints once the first prototype results are in.
The Core, Sense and Learn design, the two-model structure and the design rules, framed against prior art.
Measured results on ZIV: how fast Core clamps and stops when the AI issues unsafe commands.
Sector configurations, signed loading, and mapping the design to IEC 61508 and ISO 13849.
Floorplan, lockstep and monitors, and the limits of mask ROM against physical attacks.
We are building the platform and the test rig. If you bring formal methods, security or certification expertise, we would like to talk.