⚡ electrical engineering

Synthesis of Deep Neural Networks with Safe Robust Adaptive Control for Reliable Operation of Wheeled Mobile Robots

This paper presents a hierarchical control framework for heavy-duty wheeled mobile robots that integrates a high-precision deep neural network with a model-free robust adaptive controller and dual safety layers to ensure stable, reliable operation under disturbances while adhering to safety standards, as validated by real-time experiments on a 6,000 kg robot.

Original authors: Mehdi Heydari Shahna, Jouni Mattila

Published 2026-04-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mehdi Heydari Shahna, Jouni Mattila

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Teaching a Heavy Robot to Drive Safely

Imagine you have a massive, 6-ton construction robot (like a giant, autonomous bulldozer) that needs to drive around a construction site. This robot is heavy, powerful, and operates in messy, unpredictable environments with mud, rocks, and sudden bumps.

The researchers faced a classic dilemma: How do you make this robot drive perfectly fast and precise, while also making sure it never crashes or breaks if something goes wrong?

They tried two different approaches, but neither was perfect on its own. So, they built a three-layer safety system that switches between them like a smart driver changing gears.

The Two "Drivers" (Control Policies)

To understand the solution, you first need to understand the two "drivers" the robot uses:

1. The "Genius Student" (The Deep Neural Network / DNN)

What it is: A computer brain trained on millions of data points. It has "memorized" how the robot moves.
The Analogy: Think of this as a genius driving student who has practiced on a perfect, empty track for years. They know exactly how much to turn the wheel to hit a curve perfectly. They are incredibly fast, precise, and smooth.
The Problem: This student is a "black box." We don't fully understand how they think. If the student encounters a situation they've never seen before (like a sudden, massive rock or a slippery patch of ice), they might panic or make a wild guess because they haven't seen that specific scenario in their training data. They are great at following the rules but bad at handling chaos.

2. The "Grumpy Veteran" (The Robust Adaptive Control / RAC)

What it is: A traditional, math-heavy control system designed specifically to handle chaos.
The Analogy: Think of this as an old-school, grumpy veteran driver. They aren't as smooth or fast as the student. They might drive a bit roughly, overshoot turns, or take a wider path. But, if a rock flies at the windshield or the road disappears, this veteran knows exactly how to stabilize the vehicle using pure physics and experience. They are slow and clunky, but they are unbreakable.
The Problem: If you let the veteran drive the whole time, the robot would be too slow and inaccurate for delicate tasks.

The Solution: The "Hierarchical Safety Team"

The researchers realized they didn't have to choose between the Genius Student and the Grumpy Veteran. Instead, they created a hierarchical control policy with two safety layers (like a manager and a security guard) to switch between them automatically.

Here is how the system works in real-time:

Level 1: The "Genius" is in Charge (Normal Conditions)

The Scenario: The robot is driving on a smooth, normal path.
The Action: The Genius Student (DNN) is driving. The robot moves with high precision, following the path perfectly.
The Safety Check: A Low-Level Safety Monitor is watching the Student's performance. It's like a supervisor checking if the student is staying within the lane lines.

Level 2: The "Switch" (When Trouble Starts)

The Scenario: Suddenly, a heavy disturbance hits the robot (e.g., a huge bump, a wheel slips, or the ground is uneven). The Student starts to drift or lose control.
The Action: The Low-Level Safety Monitor sees the error getting too big. It immediately pulls the Student out of the driver's seat and locks the door (so the Student can't take over again).
The Switch: It hands the wheel to the Grumpy Veteran (RAC).
The Result: The robot's driving style changes. It becomes a bit rougher and less precise, but it instantly stops drifting and stabilizes. The robot is now "safe but slow." It can keep working through the disturbance without crashing.

Level 3: The "Emergency Stop" (When Things Get Too Bad)

The Scenario: The disturbance is so massive (e.g., the robot is about to tip over or fall into a pit) that even the Grumpy Veteran can't handle it. The robot is in genuine danger.
The Action: A High-Level Safety Monitor (the ultimate boss) sees that the Veteran is struggling and the robot is about to break.
The Result: The High-Level Monitor hits the Emergency Stop button. The robot shuts down immediately to prevent damage to itself, the environment, or people nearby.

Why This Matters

This paper is important because it solves a major problem in robotics: Trust.

Old Way: We used to rely on complex math models to drive robots. But for heavy, messy machines, the math is too hard to get right.
New Way (DNN only): We tried using AI (the Student) because it's smart and fast. But we were scared to use it on heavy machines because if it fails, it could be catastrophic.
This Paper's Way: We use the AI for speed and precision, but we wrap it in a "safety cage" (the Veteran and the monitors). If the AI gets confused, the safety cage takes over.

The Real-World Test

The researchers tested this on a real 6,000 kg (13,000 lbs) robot.

They drove it in a circle using just the AI (The Student). It was perfect.
They drove it using just the Veteran. It was safe but a bit wobbly.
The Final Test: They drove it, and then intentionally messed with the controls to simulate a huge disturbance.
- The AI tried to handle it but started to fail.
- The system instantly switched to the Veteran.
- The robot kept driving safely, even though the path wasn't perfect.
- The system proved it could switch back and forth (or rather, switch once and stay safe) without crashing.

The Takeaway

This research paves the way for using "black box" AI in heavy industry. It says: "We can use the smart, fast AI, as long as we have a safety net that knows exactly when to step in and take control." It's like having a self-driving car that knows when to call a professional driver to take the wheel if the road gets too dangerous.

1. Problem Statement

The paper addresses the challenge of deploying Deep Neural Networks (DNNs) for the control of heavy-duty Wheeled Mobile Robots (WMRs) (e.g., construction machinery weighing ~6,000 kg). While DNNs offer high-precision control without the need for complex, computationally expensive dynamic modeling, they present significant safety risks:

Black-box Nature: DNNs lack interpretability, making formal stability analysis difficult.
Lack of Robustness: DNNs trained on specific data distributions often fail or become unstable when exposed to external disturbances (e.g., terrain changes, slippage) or faults outside their training set.
Safety Standards: Heavy-duty robots must adhere to strict international safety standards (ISO 3691, ISO 13849, IEC 61508). Pure learning-based approaches struggle to guarantee safety in unpredictable, harsh environments.

The core question is: How can high-performance learning-based control be deployed on heavy-duty robots while guaranteeing stability and safety under severe disturbances?

2. Methodology

The authors propose a hierarchical control framework that synthesizes a high-accuracy DNN policy with a model-free Robust Adaptive Control (RAC) policy, monitored by two distinct safety layers with different levels of authority.

A. System Modeling

Platform: A PMSM-powered Skid-Steer Wheeled Mobile Robot (SSWMR) with a multi-stage actuation chain (Electric Motor $\to$ Hydraulic Pump $\to$ Hydraulic Motors $\to$ Gearbox $\to$ Wheels).
Model-Free Approach: The system is treated as a "black box." The control design does not rely on explicit knowledge of the complex actuator dynamics or wheel-ground interactions. Instead, it assumes bounded unknown nonlinear functions representing disturbances.

B. Control Policies

Primary Controller (LM-trained DNN):
- Training: A feedforward DNN is trained using the Levenberg–Marquardt (LM) algorithm on 1,000,000 input-output data pairs (wheel velocity vs. motor RPM) collected from the robot.
- Function: Maps reference velocities to control commands (RPM) with high precision under nominal conditions.
- Limitation: High accuracy but low robustness to unknown disturbances.
Backup Controller (Model-Free Robust Adaptive Control - RAC):
- Design: A stability-guaranteed adaptive controller based on Prescribed Performance Control (PPC) and a Logarithmic Barrier Lyapunov Function (BLF).
- Function: Compensates for external disturbances and uncertainties without requiring a system model.
- Trade-off: It ensures stability and robustness but at the cost of reduced tracking accuracy compared to the DNN.

C. Hierarchical Safety Architecture

The system employs two safety layers to manage the transition between controllers:

Low-Level Safety Layer (Performance Monitoring):
- Monitors the tracking error of the DNN controller against a tight, time-varying bound ( $\zeta(t)$ ).
- Mechanism: Uses a latched-off logic. If the DNN error exceeds $\zeta(t)$ , the DNN is permanently deactivated, and the RAC is activated. This switch is irreversible for the current session to prevent "chattering" between controllers.
High-Level Safety Layer (System Integrity):
- Monitors the system (whether under DNN or RAC) against a wider, more relaxed bound ( $o(t)$ ).
- Mechanism: If the disturbance is so severe that even the RAC cannot keep the error within $o(t)$ , the system triggers a controlled shutdown to prevent damage to the robot or environment.

3. Key Contributions

Hierarchical Synthesis: A novel architecture combining a high-precision DNN with a robust RAC, managed by a latched switching mechanism. This allows the system to enjoy high accuracy during normal operation while retaining a fail-safe mechanism for disturbances.
Model-Free Robustness: The RAC strategy guarantees uniform exponential stability for the entire system (including the switching time) without requiring explicit modeling of the complex hydraulic/mechanical actuation chains.
Compliance with Safety Standards: The framework aligns with ISO/IEC TR 5469 guidelines for AI functional safety, using supervisory barrier functions to detect unsafe behavior and initiate fallback strategies.
Adjustable Performance Bounds: The system allows for tunable safety thresholds (overshoot and steady-state error) at both safety layers, adapting to specific application requirements.

4. Experimental Results

The proposed framework was validated on a 6,000 kg heavy-duty SSWMR (Multi-Purpose Deployer) through three experiments:

Experiment 1 (DNN Only): The trained DNN achieved high-precision circular path tracking with minimal error (MSE $\approx 1.5 \times 10^{-3}$ ) under nominal conditions. However, it lacked inherent robustness to severe disturbances.
Experiment 2 (RAC Only): The RAC policy successfully maintained stability under external disturbances (simulated by 8% input increases) but exhibited larger tracking errors and transient overshoot compared to the DNN, confirming the robustness-accuracy trade-off.
Experiment 3 (Synthesis):
- The robot followed an S-shaped path.
- At $t=82$ seconds, a 15% external disturbance was introduced.
- The Low-Level Safety Layer detected the DNN error exceeding the threshold.
- The system latched to the RAC policy.
- Result: The robot continued operating safely. While tracking accuracy decreased slightly (as expected with RAC), the system remained stable and within the high-level safety bounds. The transition was seamless, and no shutdown was triggered.

5. Significance

Bridging the Gap: This work provides a practical pathway for deploying "black-box" AI controllers in safety-critical, heavy-duty industrial robotics, a domain previously dominated by conservative model-based control.
Safety Assurance: By integrating formal stability proofs (Lyapunov-based) with data-driven learning, the paper demonstrates that learning-based control can meet rigorous international safety standards (ISO/IEC).
Operational Continuity: The latched switching mechanism ensures that once a disturbance is detected, the system prioritizes stability over accuracy, allowing the robot to complete a task or reach a safe state rather than failing immediately.
Scalability: The model-free nature of the RAC and the data-driven nature of the DNN make this approach applicable to various heavy-duty robots with complex, hard-to-model actuation systems.

In conclusion, the paper successfully demonstrates that a hybrid control strategy can leverage the precision of Deep Learning while mitigating its risks through a mathematically guaranteed, robust fallback system, enabling reliable autonomous operation in harsh, unstructured environments.