Lunar Lander Hazard Avoidance
Landing on the Moon requires extreme precision and the ability to autonomously identify and avoid hazards like craters, rocks, and uneven terrain in real-time. This challenge tasks developers with creating a robust, AI-powered system for an Argonaut-like lunar lander that can process multi-sensor data (e.g., LIDAR point clouds, high-resolution optical imagery) during the critical final descent phase. The goal is to detect potential landing site hazards and recommend safe alternative landing spots or initiate avoidance maneuvers. Participants will build a system that fuses data from simulated onboard sensors, applies computer vision and spatial analysis techniques to build a terrain map, identifies unsafe regions, and computes a safe landing trajectory or target. The solution must prioritize real-time performance and reliability, as typical lunar landing windows are very narrow. Leveraging modern deep learning frameworks and efficient deployment strategies will be key to success in this mission-critical scenario.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop an autonomous lunar lander hazard avoidance system using Phi-3 for constraint understanding, TorchServe for real-time vision model deployment, and Ludwig for efficient model development, fusing LIDAR and optical data.
How work is evaluated
The evaluation will simulate a lunar lander's final descent. The system will receive streaming sensor data and must output identified hazards and a safe landing zone/trajectory. Performance will be judged on hazard detection accuracy, real-time processing capabilities, and successful avoidance.
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Implement a sensor fusion pipeline to combine simulated LIDAR point cloud data and optical camera images into a unified, navigable terrain map, generating elevation models and surface reflectivity maps.
Design and train a deep learning model (e.g., semantic segmentation, object detection) using Ludwig for declarative model definition to identify lunar hazards (craters, boulders, slopes) from processed sensor data.
Optimize the trained hazard detection model for real-time inference on resource-constrained hardware, applying techniques like quantization or pruning.
Deploy the optimized hazard detection model as a low-latency inference service using TorchServe, ensuring it can process incoming sensor frames at the required frame rate for autonomous navigation.
Integrate Phi-3 to process natural language descriptions of mission-specific constraints (e.g., "avoid shadows greater than X meters," "prioritize landing zones with slope less than Y degrees") and inform the autonomous landing site selection logic.
Build an autonomous decision-making module that uses the hazard map to identify potential safe landing sites and, if necessary, compute a feasible trajectory adjustment or alternative landing target while respecting fuel and time constraints.
Evaluate the system's performance in a simulated lunar descent environment, measuring detection accuracy, false positive rate, and the success rate of safe landing zone identification within strict time limits.
Reference links and supporting material
Sample data for 3 tasks
Challenge setup
The evaluation will simulate a lunar lander's final descent. The system will receive streaming sensor data and must output identified hazards and a safe landing zone/trajectory. Performance will be judged on hazard de...
Challenge input
Dataset of simulated LIDAR point clouds and optical images with labeled hazards (e.g., training_data/lidar/*.ply, training_data/optical/*.png, labe...
Agent execution
No agent tool is listed for this challenge.
Evaluated output
Serialized PyTorch model (e.g., .pt or .pth) trained with Ludwig.
- Verify the submitted model can be loaded by TorchServe.
- Ensure the TorchServe deployment starts successfully.
- Check if the hazard detection model can process sensor frames at a minimum required rate (e.g., 10 Hz).
- HazardDetectionF1Score target: 0.9
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Verify the submitted model can be loaded by TorchServe.
- Ensure the TorchServe deployment starts successfully.
- Check if the hazard detection model can process sensor frames at a minimum required rate (e.g., 10 Hz).
Metrics
- HazardDetectionF1Score · Target 0.9
- SafeLandingSuccessRate · Target 0.85
- DecisionLatency · Target 50
Workflow tasks
- Task 1
Input: Dataset of simulated LIDAR point clouds and optical images with labeled hazards (e.g., training_data/lidar/*.ply, training_data/optical/*.png, labels.csv).
Output: Serialized PyTorch model (e.g., .pt or .pth) trained with Ludwig.
Execution harness: Python
- Task 2
Input: A directory containing the serialized model and any custom handlers.
Output: TorchServe model archive (.mar) and deployment manifest (e.g., config.properties).
Execution harness: Docker
- Task 3
Input: Python script that receives processed sensor data (e.g., terrain map, hazard detections) and a natural language mission constraint string, and outputs a recommended safe landing coordinate.
Output: JSON object: {"safe_latitude": float, "safe_longitude": float, "avoidance_maneuver_suggested": bool}
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.