- The paper demonstrates that integrating monocular depth estimation with edge AI overcomes stereo vision limitations for planetary rover navigation.
- It employs a hybrid strategy pairing asynchronous UniDepthV2 with fast YOLO12n detection, balancing latency and real-time obstacle response.
- The study benchmarks key metrics such as MAE, processing speed, and thermal performance, offering actionable insights for low-power, real-world deployments.
Depth-Aware Rover: Monocular Metric Depth Estimation and Edge AI for Low-Cost Planetary Navigation
Introduction
This paper presents a comparative evaluation of stereo vision and monocular deep learning-based depth estimation methods for autonomous rover navigation, with a focus on practical deployment using edge AI. The primary objective is to bridge the gap between simulation and real-world application, especially under hardware and resource constraints inherent to planetary rover systems. Previous reliance on stereo vision for planetary navigation is critically assessed, and the viability of monocular approaches, leveraging recent advances in metric depth estimation and real-time object detection, is thoroughly explored.
Methodology
Simulation Environment
A photorealistic lunar surface environment was constructed in Unity, utilizing a stereo camera system to generate synthetic depth data. Disparity maps were computed using OpenCV’s StereoSGBM algorithm. The virtual rover, based on the Espacial Explorer T-30 design, executed navigation tasks with semantic obstacle detection provided by a custom-trained YOLO12n model, deployed via the ONNX backend within Unity. The system allowed for halting execution upon the detection of hazardous obstacles, mirroring operational protocols in planetary robotics.
Physical Rover Implementation
The physical prototype utilized a Raspberry Pi 4B (4GB RAM) as the compute platform, with a Logitech C270 webcam as the singular vision sensor. The drive system comprised dual DC motors interfaced via an L298N H-bridge, allowing for PWM-based velocity control. Monocular metric depth estimation was realized using UniDepthV2 in ONNX format, running asynchronously (latency ≈ 7 s/frame), while YOLO12n performed real-time object detection at 10 FPS. The software stack employed Python 3.11 with a modular design approach, encompassing depth estimation, object detection, and motor control subsystems. A Tkinter-based GUI supported manual override, live video feeds, and annotated overlays. System communication relied on Bluetooth networking for low-latency VNC access.
Evaluation Metrics
Four key metrics governed the evaluation:
- Depth Accuracy: Mean Absolute Error (MAE) between predicted and ground-truth distances (0.15 m to 2.0 m).
- Processing Performance: Frame rate, latency, CPU load, and thermal stability.
- Navigation Effectiveness: Obstacle course completion rate, elapsed time, and path fidelity.
- System Robustness: Stability under sustained operation and environmental perturbations.
Results
Simulation vs. Real-World
Stereo vision in simulation produced accurate and dense depth maps, aligning with expectations due to idealized conditions. However, real-world testing highlighted significant degradation in stereo performance, mainly due to sensitivity to camera calibration and external disturbances (e.g., lighting changes, vibration-induced misalignment).
Transitioning to monocular depth estimation with UniDepthV2 addressed these calibration and robustness issues. The ONNX-optimized UniDepthV2 maintained adequate MAE across the tested range but with substantial latency (≈ 7 s/frame), precluding real-time operation on a Raspberry Pi 4B. Nonetheless, decoupling depth and detection, with asynchronous metric depth runs and high-frequency YOLO12n inference (10 FPS, NCNN backend), enabled the system to achieve safe navigation and timely obstacle response.
System-Level Observations
- Inference Speed: YOLO12n (NCNN) operated at 10 FPS compared to 1 FPS for ONNX; UniDepthV2 (ONNX) delivered depth at ~0.1 FPS.
- Thermal Stability: System temperatures remained within 40°C–65°C under load, attributed to input resolution adjustment and model quantization.
- Robustness: The system maintained performance across repeated trials and environmental variations, with monocular methods displaying higher operational versatility.
- Hybrid Perception Strategy: The study empirically validated a hybrid scheduling approach pairing fast object detection with slower depth estimation, supporting safe autonomous navigation on low-power COTS hardware.
Key contradictory finding: While stereo is typically favored for precise metric depth in academic literature, this study’s real-world results contradict this bias, showing that low-cost monocular approaches can be more practical and robust for edge-deployed systems in unpredictable field conditions.
Implications and Future Directions
This research validates that modern monocular metric depth estimation, combined with state-of-the-art lightweight object detection, can support practical autonomous navigation on severely resource-limited platforms. Implications include:
- Practical Scalability: Monocular approaches substantially reduce hardware cost and complexity, critical for large-scale or swarms of exploratory rovers.
- Simulation-Deployment Gap: The strong simulation performance of stereo vision must be tempered by the reality of real-world operational fragility; thus, evaluation protocols must include field deployment on target hardware.
- System Optimization: Deployment on edge hardware necessitates aggressive model optimization (quantization, scheduling, heterogeneous backends) and thermal management.
The paper indicates future integration of sensor fusion (e.g., monocular-IMU/stereo fallback), digital twin frameworks for simulation-to-reality synchronization, and field trials in lunar analog environments as vital next steps. This trajectory aligns with the increasing trend toward lightweight, adaptive perception stacks for off-world robotics.
Conclusion
This study provides robust evidence that monocular deep learning-based metric depth estimation, combined with lightweight real-time object detection, is not only viable but preferable to traditional stereo methods for deployment on edge AI platforms in planetary navigation scenarios. The detailed quantitative and system-level analysis—especially latency and thermal management—offers crucial design insights for the next generation of autonomous rovers. The hybrid perception architecture and system optimization strategies demonstrated here inform both theoretical and applied advances in autonomous exploration robotics.
