A research platform designed to implement SATA trust evaluation, HMAA mission authority control, and CARA recovery enforcement for resilient autonomous systems. The testbed is designed to enable controlled experiments on autonomy degradation, sensor trust evaluation, and mission recovery under contested operating conditions. The broader research program also extends these governance principles to manipulation systems, including authority-controlled robotic arms.
This platform is a physical research testbed whose build is in progress: design and bill of materials complete; procurement and assembly in progress; no physical test data yet (as of October 2026). Its governance software is demonstrated in simulation, and all scenarios are simulation-only. This testbed is part of a unified research framework for authority-governed autonomous systems spanning single-agent, multi-agent, and manipulation domains.
Design Complete · Implementation In ProgressZenodo Publication: Oktenli, B. (2026). Authority-Governed Assured Autonomy Rover Testbed: System Architecture, Governance Design, and Reproducible Artifact Package (v1.0). Zenodo. https://doi.org/10.5281/zenodo.19143190
Autonomous systems increasingly operate in environments where sensor deception, communication disruption, and adversarial interference threaten mission safety. The autonomy architectures reviewed for this work typically rely on binary fail-safe triggers or simple threshold-based switching, without formal mechanisms for graded authority degradation and structured recovery.
Authority-governed autonomy addresses this gap by making trust assessment, authority computation, and recovery enforcement first-class components of the autonomy stack. Rather than asking "Is the AI correct?", authority-governed systems ask: "Under what constraints should the AI be allowed to act given current trust conditions?"
Research in resilient autonomy and trusted AI has been identified as a strategic priority by organizations including DARPA (Assured Autonomy program), NASA (autonomous planetary rover operations), and the National Science Foundation. DoD Directive 3000.09 (January 25, 2023) requires that autonomous and semi-autonomous weapon systems be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force.
The Authority-Governed Autonomy Testbed is designed to experimentally evaluate governance architectures for autonomous systems operating in safety-critical and adversarial environments.
The platform allows controlled experiments involving:
The testbed is designed to provide a practical evaluation environment for the SATA, HMAA, and CARA architectures described in the technical reports published on Zenodo (the architectures are also the subject of U.S. provisional patent applications, which are not published). Unlike purely simulated environments, this experimental research platform is designed to validate governance behavior with physical sensors, real-time computation, and hardware safety interlocks.
The testbed design specifies a complete authority-governed autonomy pipeline, demonstrated in simulation. Every autonomous decision must pass through trust evaluation, authority computation, and recovery logic before reaching actuators.
Beyond rover navigation, the same authority-governed autonomy framework provides a generalizable governance layer for robotic manipulation systems performing safety-critical tasks. In this extension, manipulator actions are not executed directly from perception and planning outputs; they are gated through trust evaluation and authority governance before motion commands reach the robotic arm controller.
This manipulation-oriented extension is currently under development as a second application domain for the broader authority-governed autonomy research program.
This extension, currently under development, is intended to show that authority-governed autonomy is not limited to navigation systems, but can govern fine-grained manipulation tasks where incorrect actions may cause physical damage or safety risks. This highlights the framework's intended applicability to a broad class of autonomous systems operating under uncertainty.
The testbed is also designed as a reusable sensor trust testing platform to evaluate trust-aware autonomy systems under controlled adversarial and degraded conditions. This platform is designed to enable systematic injection of sensor faults to study trust degradation, disagreement detection, and authority response.
Multi-sensor array (camera, LiDAR, IMU, ToF), programmable fault injection mechanisms, embedded compute for trust evaluation, real-time telemetry and logging
Sensor spoofing detection and response, sensor degradation modeling, cross-sensor disagreement analysis, trust-based authority adjustment validation
This platform is intended to support experimental research on resilient perception systems and to provide a controlled environment for testing trust evaluation algorithms under adversarial conditions.
Together with the rover testbed and robotic manipulation extension, this platform is designed to demonstrate a unified authority-governed autonomy framework spanning perception, decision-making, and actuation layers.
Full electrical schematic showing MCU, sensor, actuator, power, and module interconnections. Color-coded by node type: blue (MCU), teal (Sensor), orange (Actuator), yellow (Power), green (Module).
The rover testbed simulation environment provides a controlled experimental platform for evaluating authority-governed autonomy under adversarial and degraded sensor conditions. The simulator executes the complete SATA trust fusion, HMAA authority computation, and CARA recovery logic in real-time with configurable fault injection.
This simulation is an executable demonstration of authority-governed autonomy, not only a conceptual design. The simulation environment mirrors the rover hardware design, so that governance behaviors demonstrated in simulation can be carried over to hardware testing once the build is complete; no physical test data exist yet.
Camera occlusion, LiDAR spoofing, IMU drift, RF disruption, compound failures
Real-time fused trust score, per-sensor values, Dempster-Shafer fusion, decay/recovery
Authority state transitions, command gating, CARA recovery activation, mission metrics
Experimental Simulation Environment (Research Use). This environment is the primary demonstration layer for authority transitions and recovery behaviors, supporting repeatable experimentation before physical hardware testing.
Launch Governance SimulatorThe testbed design uses a four-level authority state machine with hysteresis guards, dwell timers, and oscillation prevention. Upgrade thresholds are set higher than downgrade thresholds to ensure the system does not prematurely restore autonomy after trust degradation.
All sensors trusted. Autonomous navigation, planning, and mission execution. No operator input required.
Partial trust degradation. Planner restricted to conservative paths. Speed limits enforced. Operator alerted.
Significant trust loss. Loiter or return-to-base mode. Operator supervision required for any action beyond safe hold.
Critical trust failure or watchdog timeout. All motors disabled. System awaits operator recovery command or physical reset.
Transition rules include configurable dwell timers (minimum time at each level before upgrade), asymmetric thresholds (downgrade at τ < 0.6, upgrade only at τ > 0.8), and an oscillation guard preventing more than 2 transitions within any 10-second window.
The trust evaluation subsystem computes a continuous trust scalar from multi-sensor fusion using a weighted Dempster-Shafer belief model with cross-sensor validation:
Trust(s_i) = weighted belief function with cross-sensor consistency checks, disagreement penalties, asymmetric decay (fast) and recovery (slow), and single-sensor veto capability
Key properties: trust decays rapidly when sensors disagree but recovers slowly when consistency returns. A single sensor producing readings inconsistent with all others triggers an immediate trust penalty regardless of its historical reliability. This asymmetry prevents the system from trusting potentially spoofed inputs too quickly after an anomaly.
The testbed uses a dual-compute architecture separating high-level autonomy (Raspberry Pi 5) from real-time safety control (ESP32), connected via UART with hardware time synchronization.
Raspberry Pi 5 8GB: ROS 2 nodes, SATA trust, HMAA authority, CARA recovery, mission planner, sensor fusion
ESP32-DevKitC-32D: motor PWM, encoder feedback, watchdog heartbeat, LoRa comms, command gate enforcement
RPLIDAR A1M8 (USB), VL53L1X ToF (I2C), BNO085 IMU (I2C), RPi Camera v3 (CSI), magnetic wheel encoders (GPIO)
TPL5110 watchdog, mushroom E-stop, ATECC608B secure element, RFM95W LoRa 900MHz, MicroSD audit logger
| Component | Model | Purpose |
|---|---|---|
| Main Compute | Raspberry Pi 5 (8GB) | Autonomy stack, sensor fusion, ROS 2 |
| Safety Controller | ESP32-DevKitC-32D | Real-time control, watchdog, command gate |
| LiDAR | Slamtec RPLIDAR A1M8 | 360-degree obstacle sensing |
| IMU | Adafruit BNO085 9-DOF | Orientation and motion fusion |
| Camera | Raspberry Pi Camera v3 | Visual perception (Sony IMX708) |
| ToF Sensor | Adafruit VL53L1X | Close-range obstacle detection |
| Communication | Adafruit RFM95W (900 MHz) | LoRa long-range telemetry |
| Secure Element | ATECC608B TrustFLEX | Cryptographic key storage |
| Motor Driver | Pololu Dual G2 18v18 | Dual high-power motor control (18A/ch) |
| Drive Motors | Pololu 12V 75:1 Gearmotor (x2) | High-torque differential drive |
| Safety | E-stop + TPL5110 Watchdog | Hardware safety interlock layer |
| Power | 3S 11.1V 5000mAh LiPo + 10A UBEC | Mobile platform power with regulation |
37 components specified, with sources checked. All commercially available. Full BOM available as downloadable CSV.
Seven experiments designed to validate the authority-governed autonomy thesis. Each experiment injects a specific fault condition and measures authority response, mission behavior change, and recovery activation.
Each experiment requires minimum 30 trials for statistical significance. Metrics collected: trust scores, authority transitions, recovery activation events, mission success rate, unsafe action count (baseline vs. governed).
Status (as of October 2026): design and bill of materials complete; procurement and assembly in progress; no physical test data yet. The architecture, authority model, trust evaluation method, and experimental methodology are defined and documented.
Complete engineering documentation for the Authority-Governed Autonomy Testbed. All files are original work by Burak Oktenli.
This project provides open artifacts (source, tests, TLA+ specification, seeded generator and simulator) for independent examination of the system architecture. Independent replication of the governance experiments is not claimed: the deposit does not currently expose a run-level result or trace archive, and legacy aggregate run-count claims are withdrawn. All documentation, schematics, and specifications are available for download.
Complete blueprint PDF, electrical schematic SVG, wiring connections JSON, system configuration JSON
37-component BOM with verified sources. All commercially available. Total cost under $500. Assembly guide included.
7 defined experiments with fault injection procedures, expected results, and statistical requirements (30+ trials each).
Fused trust scores, per-sensor trust values, authority state transitions, recovery latency, mission success rate.
The rover testbed design specifies an 11.1V 3S LiPo battery (per the bill of materials) with 5V and 3.3V regulated rails for compute and sensor modules.
11.1V 3S LiPo battery → 10A 5V UBEC (RPi5, ESP32 and sensors); 3.3V rails from the RPi5 and ESP32 boards. The motor driver is fed through the E-stop and a step-down regulator.
Emergency stop: a normally-closed mushroom E-stop in the motor supply path cuts motor power directly; separately, the ESP32 heartbeat to the TPL5110 watchdog gates the motor-driver enable line, which is designed to disable the motors on a missed heartbeat. RPi5 and ESP32 remain powered for governance state preservation and audit logging.
The testbed design specifies a simplified security chain appropriate for a research platform while maintaining governance integrity.
A TLC model check of the HMAA discrete authority model (HMAA_Authority_FSM, archived at Zenodo 10.5281/zenodo.21300977) at small constants (MaxTick = 8, MaxDwell = 8, OscMaxTransitions = 3, OscLockoutTicks = 4) explored 23,748 distinct states to depth 9 with no counterexample. Of the eight properties checked, four are exercised and four hold only vacuously at these bounds because no upgrade and no lockout is reachable. The exercised properties are TypeInvariant, LevelBounded, DwellNonNegative and CompoundFailureReachesA0; the vacuous ones are UpgradeIsStepwise, TrustAuthorityWeakConsistency, LockoutConsistent and OscillationLockoutTerminates. The run at operational constants did not complete. The model assumes instantaneous authority equals its target (continuous behavior between decision instants is not covered) and numbers authority levels inversely (level 0 is full authority). This is a bounded check of the HMAA authority model only, not of the SATA-HMAA-CARA pipeline or the code.
42-file Python engineering baseline with 98 unit tests and 200,000 FSM conformance comparisons.
Complete governance state log with timestamps. Every authority transition recorded for post-mission analysis.
| Subsystem | Cost | % of Total |
|---|---|---|
| Compute (Raspberry Pi 5 + ESP32-S3) | $95 | 20% |
| Sensors (ultrasonic + IMU + camera) | $65 | 13% |
| Drive System (motors + driver + wheels) | $120 | 25% |
| Chassis & Structure | $85 | 18% |
| Power (battery + regulators) | $45 | 9% |
| Wiring, connectors, misc | $74 | 15% |
Low-cost governance testbed design for the complete SATA-HMAA-CARA pipeline at ~$484. All COTS components.
| Parameter | Value |
|---|---|
| Components | 37 (dual-compute: RPi5 + ESP32-DevKitC-32D) |
| Simulation runs | seven fault scenarios demonstrated in simulation |
| Formal model check | TLA+ bounded model check of the HMAA authority model (23,748 distinct states, depth 9, no counterexample; 4 of 8 properties exercised, 4 vacuous at these bounds; archived at Zenodo 10.5281/zenodo.21300977) |
| Test suite | 98 tests, 200,000 FSM conformance comparisons |
| Governance pipeline | SATA → HMAA → CARA (8-stage) |
| Power | 11.1V 3S LiPo (per the bill of materials) |
| BOM | ~$484 |
Hardware testbeds for supervisory safety architectures have an established history. The Simplex line has been demonstrated on real-time embedded platforms since the late 2000s, including system-level implementations placing the supervisor on separate hardware, and later work extended those demonstrations to unmanned aerial platforms and to learning-enabled controllers. Dedicated safety supervisors running on independent hardware are also routine in industrial robotics (ISO 10218), in process-industry safety instrumented systems (IEC 61511) and in automotive functional safety practice (ISO 26262). This testbed claims no novelty in the idea of validating a supervisory architecture on physical hardware.
Its stated purpose is narrower and should be read as such: to exercise an AUTHORITY decision, meaning whether a system is currently permitted to act given trust, integrity and context, on real actuation with real sensor degradation, and to measure recovery behaviour when authority is restored incrementally rather than switched back. The comparison it is designed to support is against a matched configuration lacking the authority layer, not against the runtime-assurance literature as a whole. Where the prior testbeds validated that a supervisor can take control safely, this one is designed to observe what happens on the way back.
This testbed represents the first phase of a broader research program in authority-governed autonomy. Future research directions include:
Authority coordination across multiple autonomous agents with distributed trust aggregation (MAIVA integration)
Integration of ADARA deception probability engine for proactive authority adjustment under sensor manipulation
Field experiments in outdoor contested environments with GPS spoofing, communication jamming, and multi-sensor failure injection
FLAME deliberation windows applied to physical actuator command chains with measured response timing
Extension of the governance framework to robotic manipulation, where perception uncertainty dynamically constrains motion authority, action permissions, and manipulator behavior
This project is part of ongoing research on authority-governed autonomy and resilient AI systems conducted independently by Burak Oktenli. The testbed design uses the same governance architectures described in the SATA, HMAA and CARA technical reports on Zenodo (the architectures are also the subject of U.S. provisional patent applications, which are not published). Current work focuses on the rover testbed as the first physical platform (build in progress), with planned extensions to robotic manipulation systems and additional authority-governed autonomy domains.
Related research architectures: SATA (sensor trust), HMAA (authority computation), CARA (recovery), MAIVA (multi-agent trust), FLAME (latency control), ADARA (deception-aware risk).