Autonomous exploration
Acheron provides autonomous navigation through multi-layered planning systems and maintains spatial awareness through perception modules.
Spatial memory
Creates semantic maps of the environment by linking visual data with spatial locations. The system continuously records video frames and odometry, storing only key frames when the robot moves beyond a set distance or time threshold.
Vector database and embedings
Visual data is transformed into vector embeddings using CLIP and ResNet models. These embeddings are stored in a vector database optimized for fast search. The system supports three search methods: text queries (natural language search), location queries (search within a spatial area), and image queries (finding visually similar memories).
Object and human detection
Our tracking system supports both generic object tracking and dedicated person tracking. For objects, it combines advanced techniques like CSRT tracking, ORB re-identification, 3D distance estimation, and depth integration. For people, it uses a YOLO-based pipeline that detects humans, filters them from other objects, estimates 3D positions, and visualizes results with bounding boxes.
The spatial memory system integrates with navigation to enable natural language commands. Navigation module uses a two-step approach: first trying to locate objects in the current camera view, and then searching stored spatial memory when needed.