AGRO: Autonomous AI Rover

Updated 2 March 2026

AGRO is a modular, AI-enabled unmanned ground rover that autonomously navigates and performs precision tasks using advanced sensor fusion and machine learning.
It integrates robust embedded control, real-time cloud telemetry, and a hybrid deliberative-reactive planning architecture to ensure efficient obstacle avoidance and safety.
Applications span agricultural field operations and planetary exploration, validated by high-precision metrics in crop scouting, yield mapping, and terrain mapping.

AGRO (Autonomous Ground Rover Observer) is a modular, AI-enabled unmanned ground vehicle (UGV) engineered for autonomous navigation, perception, and high-precision operations in agricultural and planetary exploration domains. The AGRO framework integrates advanced machine learning, computer vision, multisensor fusion, robust embedded control, and AI-driven mission planning to automate complex field and surface tasks. The architecture and performance profile of AGRO are substantiated by results from precision agriculture deployments, simulated lunar environments, and field-validated Mars-analogue trials (Ghumman et al., 2 May 2025, Gamar et al., 26 Dec 2025, Luna et al., 7 Oct 2025, Santra et al., 27 Oct 2025).

1. Platform Architecture and Embedded System Design

AGRO platforms typically employ a 4-wheel drive or 4WIDS chassis optimized for stability and payload, using lightweight materials such as anodized aluminum and vibration-isolated ABS sensor mounts. The mechanical architecture supports up to 3 m/s speeds with payload capacities exceeding 5 kg, facilitated by high-torque, gearmotor-driven wheel assemblies. The embedded system consists of a high-level control unit (e.g., Raspberry Pi 5 with ROS 2) for perception, planning, SLAM, and cloud telemetry, paired with a real-time microcontroller (ESP32-S3 under FreeRTOS) for deterministic low-level (200 Hz) PID-based motor control and safety monitoring. Communication between control tiers is realized via low-latency (1–3 ms) UART protocols, with system-wide failsafes triggered by comms timeout or voltage anomalies (Gamar et al., 26 Dec 2025).

ROS 2 node architecture orchestrates LiDAR, RGB-D, and GNSS/IMU data, running ekf_localization (fusion-based odometry), global A* path planning, local MPPI trajectory optimization, and perception modules (YOLO3D, lane_detector). Real-time cloud monitoring and supervisory control are enabled by AWS IoT and S3 data streaming for model retraining and remote emergency interventions.

2. Perception and Sensor Fusion

The AGRO sensor suite typically integrates high-resolution RGB cameras (e.g., 64 MP Arducam), 2D/3D LiDAR (up to 350 m range), multiband RTK GNSS (sub-2 cm accuracy), high-frequency IMU, and application-specific probes (soil moisture, thermal) (Ghumman et al., 2 May 2025). Sensor data is processed by onboard computer vision pipelines for object detection (e.g., with YOLOv10/YOLOv8n, achieving up to 98.88% mAP@50 in pistachio yield estimation), semantic segmentation (DeepLabV3+, HRNet), and 3D bounding box extraction (YOLO3D + monocular depth).

Sensor fusion is accomplished via Extended Kalman Filtering, with state vectors including position $p_k$ , velocity $v_k$ , quaternion orientation $q_k$ , and inertial sensor biases. GNSS and IMU fusion achieves localization RMSE of 0.12 m. LiDAR scans are used to probabilistically update log-odds occupancy maps, enabling real-time construction of 2D/3D traversability grids (Ghumman et al., 2 May 2025, Gamar et al., 26 Dec 2025).

Table: Key Onboard Sensors and System Performance

Component	Specification / Result
RGB Camera	64 MP, 9152×6944 px
LiDAR	350 m range, 10–40 Hz
GNSS/RTK	<2 cm accuracy (RTK fixed)
Localization RMSE	0.12 m (GNSS-EKF fusion)
YOLOv10 Yield mAP@50	0.9888

Mission planning in AGRO leverages a hybrid deliberative-reactive architecture, decomposing global route selection and local obstacle avoidance. Global path planning uses A* over occupancy grids derived from fused semantic maps and vision-extracted hazards, with edge costs incorporating distance, obstacle proximity, and traversal risk (Lavin, 2015, Ghumman et al., 2 May 2025). Multi-objective cost functions can include energy, science value, and environmental risk terms.

Local planning employs both deterministic (Dijkstra, Bézier-based BendyRuler) and stochastic (MPPI) approaches for smooth, dynamically feasible trajectories maximizing clearance. AGRO also incorporates pure pursuit control and safety-constrained stopping with real-time obstacle updates from LiDAR sector masks.

Perception-augmented path planning is further enhanced by semantic segmentation (ViBEKO DeepLabV3+) and Far Obstacle Detector (FASTNAV FOD) modules, providing high-frequency (1–5 Hz) hazard rasterization and enabling rapid traverse speeds up to 1.0 m/s with safety margins under emergency conditions (Luna et al., 7 Oct 2025).

4. Machine Learning-Based Control and Policy Transfer

AGRO systems exploit transferable deep reinforcement learning (DRL) policies for robust navigation in unstructured environments. Using PPO-based DRL, agents are trained in simulated agricultural domains (IsaacSim) and evaluated in lunar-gravity scenarios. The state space incorporates 12D observations (pose, goal vectors, orientation errors, depth-based obstacle filters, raw wheel speeds), and action space issues continuous velocity commands to the base (Santra et al., 27 Oct 2025).

The DRL reward structure includes survival, goal proximity, orientation shaping, obstacle penalties, and episode-level bonuses. Policies trained on terrestrial environments achieve ~47% zero-shot goal-reaching on lunar terrain, with no collision events and stable obstacle avoidance, indicating substantial cross-domain generalization without fine-tuning.

5. Multi-Robot Coordination and System-Level Autonomy

AGRO extends to multi-agent scenarios through the CISRU framework, enabling real-time task allocation, collaborative mapping, and emergency handling among heterogeneous rovers and human operators. Coordination is achieved using DDS-backed publish–subscribe channels, MobileNet-SSD perception stacks, and hierarchical task allocation (Luna et al., 7 Oct 2025). Optimization of resource assignment is formalized as a utility maximization under robot capacity and task constraints.

Critical system-level behaviors include:

Redundant fail-safes: Automatic fallback to local autonomy if communication loss occurs, with controlled shutdown on battery or hardware anomaly.
Safety-first control: Immediate halt on emergency detection by accelerometer/vision fusion.
Prioritized communication: ROS 2 over DDS with emergency channel preemption.

Performance metrics in Mars-analogue field tests report mapping improvements (+15% coverage), sub-1.2 s emergency response, and functional area-wise task allocation.

6. Precision Agriculture Applications

In the agricultural context, AGRO automates crop scouting, yield mapping, and resource optimization through high-fidelity perception and data-driven analytics (Ghumman et al., 2 May 2025). Deployments in pistachio orchards demonstrate mission runtimes averaging 38 min per 200 m, 100% obstacle avoidance, occupancy grid IoU of 0.82 against ground truth surveys, and crop yield estimation error within ±10.66%. Modularity allows integration of multispectral and soil-moisture sensors, supporting advanced crop health modeling.

Recent works such as Agronav employ vision-based navigation pipelines—combining semantic segmentation (ViT-Adapter, HRNet) and deep Hough transform-based line detection—delivering centerline computation accuracy (mIoU up to 96.43%) and F1 line detection scores up to 0.969 in ground robotics, with real-time performance achieved by HRNet and MobileNetV3 architectures (Panda et al., 2023).

7. Limitations, Scalability, and Future Directions

Current AGRO deployments exhibit certain limitations: reliance on post-mission data offloading (planned upgrades to high-bandwidth links), occlusion-induced detection errors (to be addressed with multi-angle and transformer-based architectures), and absence of onboard real-time analytics in some prototypes (Ghumman et al., 2 May 2025). Vision-based systems face challenges with aerial images of high row multiplicity and dense weed occlusion (Panda et al., 2023). GAN-based augmentation and few-shot learning are being explored to address data scarcity for rare class detection.

Planned future work includes:

FPGA-based perception integration for <10 ms inference cycles (Gamar et al., 26 Dec 2025)
Reinforcement learning navigation stack deployment on embedded accelerators
Cross-environment policy adaptation for planetary field trials
Fully integrated, cloud-connected mission loops with real-time map streaming, remote supervision, and continuous self-supervision/retraining

AGRO represents an extensible, rigorously validated framework for autonomous AI-enabled rover operations spanning terrestrial agriculture and planetary exploration domains (Ghumman et al., 2 May 2025, Luna et al., 7 Oct 2025, Santra et al., 27 Oct 2025, Gamar et al., 26 Dec 2025).