HUMAN FIRST: A Safety Critical Brain Operating System Blueprint for Humanoid Robotics with Obligation Bounded Authority

The deployment of humanoid robots in environments shared with people demands safety guarantees that extend well beyond conventional software testing. Every motor command, every joint actuation, and every force exerted by a robotic platform must be traceable to an explicit, bounded grant of authority issued by a human operator. HUMAN FIRST was developed to provide such guarantees through a brain operating system blueprint that enforces obligation bounded authority and covenant gated execution at every layer of the control stack.

The central design question motivating this work is straightforward but consequential: how can we ensure that no autonomous motion ever occurs without verified human consent, attested hardware identity, and satisfaction of all applicable safety constraints? Existing robotic middleware frameworks provide varying degrees of safety monitoring, but few offer a formally structured mechanism that requires all three conditions to hold simultaneously before any actuator command is permitted to execute. HUMAN FIRST addresses this gap by organizing the entire control pipeline around three invariants that are enforced at the architectural level rather than through optional runtime checks.

The first invariant requires that no actuator command may proceed unless all fifteen safety covenants defined in the system specification are satisfied. These covenants cover thermal limits, force boundaries, velocity constraints, workspace boundaries, emergency stop status, and communication integrity, among others. Violation of any single covenant triggers an immediate hardware level halt through an independent safety island that operates outside the main software pipeline, ensuring that the halt mechanism cannot be overridden or delayed by faults in the cognitive layer.

The second invariant establishes that every grant of motion authority carries explicit scope, duration, and accountability metadata. This means the system can always answer three questions about any action it has taken or is about to take: who authorized it, under what constraints, and for how long that authorization remains valid. Authority grants are cryptographically signed and time bounded, preventing both unauthorized commands and stale authorizations from reaching the actuators.

The third invariant ensures that the entire control pipeline is fail closed. Any ambiguity, timeout, communication loss, or unrecognized condition results in safe cessation of motion rather than continued operation under uncertainty. This design choice reflects the position that in safety critical robotics, the cost of a false halt is almost always lower than the cost of an unauthorized or uncontrolled movement.

The implementation spans three programming languages, each chosen for the properties most relevant to its portion of the pipeline. The cognitive decision layer is written in Node.js, taking advantage of its event driven architecture for responsive decision processing. Safety verification and formal constraint evaluation are implemented in Python, where the rich ecosystem of numerical and logical libraries supports clear and auditable safety logic. Actuator boundary enforcement, where deterministic timing and memory safety are essential, is implemented in Rust.

This multi language design is not arbitrary. It reflects a deliberate engineering decision to allow each subsystem to be verified independently using the testing frameworks and static analysis tools most appropriate to its runtime characteristics. The interfaces between language boundaries are narrow and formally specified, reducing the surface area for integration defects.

The validation suite comprises 869 automated test cases distributed across the three runtimes: 799 in Node.js covering the full cognitive pipeline, 35 in Python verifying cross language safety parity, and 35 in Rust exercising actuator boundary enforcement under adversarial inputs. All tests execute with zero failures across all supported platforms.

During rigorous boundary testing, a torque limit bypass defect was identified in the actuator enforcement layer. Specifically, negative torque values were able to circumvent the magnitude comparison because the original implementation compared the signed value against the positive limit without taking the absolute value. This defect was corrected by applying absolute value comparison before the limit check, and targeted regression tests were added to verify that both positive and negative out of range values are correctly rejected.

Virtual hardware in the loop validation exercises six simulated scenarios that stress the safety architecture under realistic failure conditions. These include thermal shutdown sequences, emergency stop activation and release, communication timeout recovery, and multi joint concurrent limit violations. All six scenarios complete successfully under the fail closed protocol, confirming that the safety island responds correctly even when the cognitive layer is unresponsive or issuing contradictory commands.

Continuous integration reproduces the complete test suite across six platform and runtime combinations spanning Ubuntu, macOS, and Windows with Node.js versions 20 and 22. This cross platform verification ensures that the safety properties demonstrated in testing are not artifacts of a particular operating system or runtime version.

This release represents a software research blueprint. Physical robot integration, real time operating system adaptation for deterministic actuator timing, and formal safety certification against ISO 13482 (personal care robots) and ISO 10218 (industrial robots) remain as planned future work requiring access to production hardware platforms and accredited testing facilities.