Vision-based Goal-Reaching Control for Mobile Robots Using a Hierarchical Learning Framework

Abstract: Reinforcement learning (RL) is effective in many robotic applications, but it requires extensive exploration of the state-action space, during which behaviors can be unsafe. This significantly limits its applicability to large robots with complex actuators operating on unstable terrain. Hence, to design a safe goal-reaching control framework for large-scale robots, this paper decomposes the whole system into a set of tightly coupled functional modules. 1) A real-time visual pose estimation approach is employed to provide accurate robot states to 2) an RL motion planner for goal-reaching tasks that explicitly respects robot specifications. The RL module generates real-time smooth motion commands for the actuator system, independent of its underlying dynamic complexity. 3) In the actuation mechanism, a supervised deep learning model is trained to capture the complex dynamics of the robot and provide this model to 4) a model-based robust adaptive controller that guarantees the wheels track the RL motion commands even on slip-prone terrain. 5) Finally, to reduce human intervention, a mathematical safety supervisor monitors the robot, stops it on unsafe faults, and autonomously guides it back to a safe inspection area. The proposed framework guarantees uniform exponential stability of the actuation system and safety of the whole operation. Experiments on a 6,000 kg robot in different scenarios confirm the effectiveness of the proposed framework.

• The paper introduces a hierarchical framework that integrates vision-based SLAM, RL planning, deep actuator modeling, and adaptive control to achieve centimeter-level goal-reaching accuracy.
• It employs a modular design where visual state estimation, motion planning, and actuator dynamics modeling work together to enable smooth and safe operations under varying terrain conditions.
• Experimental results validate the approach with a 3 cm RMSE goal-reaching accuracy and superior transient response on both asphalt and loose soil, demonstrating its industrial viability.

Hierarchical Vision-Based Goal-Reaching Control for Large-Scale Mobile Robots

Introduction and Motivation

The paper "Vision-based Goal-Reaching Control for Mobile Robots Using a Hierarchical Learning Framework" (2601.00610) addresses the persistent challenge of applying reinforcement learning (RL) for goal-directed navigation and control of large, heavy-duty mobile robots operating in hazardous, partially known environments, such as mining and construction. Unlike the majority of existing RL-based methodologies, which are primarily tested on lightweight robots and often ignore the complexities of actuator dynamics, this framework provides a comprehensive, modular approach that integrates high-accuracy visual state estimation, smooth RL-based motion planning, supervised deep neural network (DNN) actuator modeling, robust adaptive low-level control, and explicit safety supervision.

Framework Overview

The proposed framework is architected as a set of tightly coupled yet functionally distinct modules:

• Visual SLAM State Estimation: Provides robust, real-time pose estimation for downstream planning and control.
• RL-Based Motion Planning: Employs Q-learning/SARSA for smooth, constraint-aware goal-reaching at the kinematic level.
• DNN-Based Actuator Modeling: Trains a DNN to capture nonlinear actuator system dynamics for each wheel.
• Robust Adaptive Control (RAC): Ensures the actuator system accurately tracks the planner's reference signals, even under terrain-induced uncertainties.
• Safety Supervisor: Monitors safety-critical conditions, enforces operational limits, and autonomously returns the robot to a safe area when necessary.

  Figure 1: The real-time interaction of all sensing, planning, actuation, and safety modules within the hierarchical framework.

RL-Based Motion Planning Module

The core planning problem is formulated as a discrete MDP where state features—robot position, heading, linear and angular velocities—are discretized for tabular RL. Notably, the design deviates from standard approaches: the agent outputs accelerations, not direct velocity or action commands, which substantially improves the smoothness of the generated motion profiles under real hardware constraints. The reward function is comprehensively engineered, combining potential-based shaping (using task distance and heading error) with additional smoothness, approach-direction, and oscillation-avoidance penalties. Deadbands and zero-lock shaping further suppress undesirable angular jitter near goal regions.

Figure 2: Block diagram of the RL-based motion planning system, including state discretization, action selection, and policy update mechanisms.

By shaping the RL reward and constraining actions contextually, the resultant policy achieves practical, monotonic, and stable robot trajectories suitable for large, inertia-dominated platforms.

Deep Neural Network Modeling of Actuator Dynamics

For each in-wheel actuator, a feedforward DNN is trained via the Scaled Conjugate Gradient (SCG) method to model the inverse dynamics—mapping desired wheel velocities to actuator control inputs. The architecture is validated using standard data splits (train/validation/test) with early stopping based on validation loss to mitigate overfitting. This component is critical for handling the highly nonlinear, slip-prone dynamics typical of large off-road robots and acts as a high-fidelity model for reference tracking.

Figure 3: Workflow of DNN training for the inverse dynamics model of the actuator system, enabling accurate velocity-to-input mapping.

Robust Adaptive Control and Safety Enforcement

The DNN actuator policy, though effective in-distribution, is insufficient under unexpected disturbances or out-of-distribution terrain conditions. An adaptive feedback law, augmented with a Lyapunov-based barrier function, guarantees uniform exponential stability for the wheel velocity tracking error and prevents unsafe divergence outside a defined operational region. The control law incorporates error-based feedback and a dynamic, logarithmic barrier on the feasible goal-reaching set, seamlessly integrating into safety supervision.

The safety supervisor monitors the robot’s zone via geometric criteria; when threshold-violating errors or faults are detected, the robot is commanded to a safe inspection area, halting unsafe operations autonomously.

Experimental Evaluation

The framework was validated on a 6,000 kg skid-steered vehicle across two disparate terrain types: asphalt and loose soft soil.

Figure 4: Test surfaces for deployment—both high-grip asphalt and challenging, loose soft-soil terrain with pronounced slip.

Notable experimental results include:

• Goal-Reaching Accuracy: Achieved $\approx 3$ cm RMSE in final goal location for both terrain types.
• Robustness Across Terrains: Similar accuracy metrics on low-slip and high-slip surfaces, corroborating the effectiveness of the adaptive controller in slip compensation.
• Safety Supervisor Effectiveness: On artificial SLAM fault injection, the robot consistently returned to the predefined safe patch, validating the supervisory logic in mixed adversarial scenarios.

  Figure 5: Quantitative results for goal-reaching performance (in meters), showing near-centimeter-level accuracy on both terrains.

  

  Figure 6: Motion tracking of reference vs. actual robot trajectories under control, demonstrating high-fidelity actuator response and minimal steady-state error.

Comparison with other robust adaptive controllers highlights superior performance in terms of reduced overshoot (0.028 m/s peak), faster rise time (4.1 s), and minimal steady-state velocity error (0.009 m/s).

Discussion and Implications

The modular decomposition adopted—in which RL-based planning (for global task efficiency) and DNN-based actuation (for local dynamic accuracy) are coupled via an explicitly engineered interface—demonstrates that safe, high-precision, vision-guided goal-reaching is attainable on large-scale mobile robots even under severe terrain nonidealities. The formal safety mechanisms and Lyapunov/barrier stability arguments move beyond the probabilistic guarantees common in perception-driven RL, providing deterministic boundaries for real-world deployment.

From a theoretical standpoint, the integration of vision-based SLAM, discrete RL, and DNN-driven inverse modeling under a robust adaptive control envelope establishes a new reference design for scaling data-driven robotics beyond research platforms to industrial-scale, safety-critical hardware.

Practically, this approach directly supports automation in domains such as mining, industrial inspection, and construction robotics, with strong implications for reducing human exposure to hazardous environments and improving operational reliability.

Future Outlook

Given the observed performance and stability, several extensions are compelling:

• Bridging the DNN actuator model to incorporate explicit uncertainty quantification for enhanced adaptation in unseen environments.
• Scaling the RL action/state space with function approximation to support richer, continuous robot kinematics.
• Online transfer learning to adapt the actuator model without full retraining when hardware configurations or soil properties drift over time.
• Integrating higher-level, multi-robot coordination and SLAM-sharing mechanisms for collaborative, fault-tolerant exploration in complex sites.

Conclusion

This work delivers a holistic, rigorously validated framework for vision-guided, safe, and robust goal-reaching control in large-scale mobile robots, synthesizing RL-based planning, DNN-driven actuation, and adaptive safety-critical control. The centimeter-level performance consistency across challenging terrains, explicit safety guarantees, and online fault recovery collectively indicate the suitability of this hierarchical learning paradigm for industrial and mission-critical field robotic applications (2601.00610).