Technical Architecture

The systems and algorithms that power Yantrax's autonomous operation.

Hardware Stack

Component Specification Purpose
Main Compute NVIDIA Jetson Orin Nano 8GB All inference, processing, autonomy logic
Perception Stereo ZED 2i Camera Depth, stereo vision, spatial awareness
Navigation LiDAR (16-channel) SLAM, obstacle detection, mapping
IMU 9-axis IMU + Magnetometer Orientation, gyroscopic stabilization
Drive System 2× DC Motors + Encoders Differential drive locomotion
Display 23" Portrait Touch Display Visitor interaction, visual feedback
Audio Stereo speakers + Microphone array Speech synthesis, multilingual guides
Power 240Wh LiPo Battery + BMS ~6 hours continuous operation

Software Stack

ROS 2 Humble

Middleware for sensor integration, module communication, and real-time constraints.

Nav2

Autonomous navigation, path planning, behavior trees for complex autonomy sequences.

TensorRT

Optimized inference engine for quantized neural networks on Jetson hardware.

OpenCV

Computer vision pipelines: feature detection, image processing, artifact analysis.

Tesseract OCR

Optical character recognition for manuscript digitization and text extraction.

Espeak-ng

Offline text-to-speech in Hindi, Tamil, Telugu, Kannada, and English.

Key Algorithms

Simultaneous Localization and Mapping (SLAM)

Yantrax uses laser-based SLAM to build real-time maps of its environment without pre-loaded floor plans. The LiDAR scans 16 layers at 10Hz, feeding into an incremental graph-based SLAM backend that maintains accuracy over hours of continuous operation.

Visitor Detection & Tracking

The stereo camera runs a lightweight person detector (MobileNet-SSD, quantized) to identify visitors and their rough distance. When a person is detected, Yantrax approaches and initiates its greeting sequence, with the display showing "Namaste" and offering language selection.

Artifact Recognition

A fine-tuned ResNet-50 model identifies museum artifacts by image classification. The model is quantized to INT8 precision using TensorRT, achieving 94% accuracy on a dataset of 2,400 Indian heritage items while running at 45 FPS on Jetson Orin Nano.

Manuscript OCR Pipeline

For digitization, Yantrax captures high-resolution photos of documents, applies adaptive thresholding and deskewing, then runs Tesseract OCR with language-specific dictionaries. The output is searchable text paired with the original image for archival.

Multilingual Natural Language

Visitor queries are processed by a small transformer model (DistilBERT + custom heads, 110M parameters) fine-tuned on heritage-related QA pairs in six Indian languages. Responses are generated as templates and synthesized to speech using Espeak-ng with prosody adjustments.

Performance Metrics

Metric Value
Max Speed 0.8 m/s
Turn Radius 0.3 m (in-place capable)
SLAM Update Rate 10 Hz
Visitor Detection Latency ~120 ms
Artifact Recognition Latency ~45 ms @ 45 FPS
Speech Synthesis Latency ~500 ms for a 30-word response
OCR Throughput ~2 pages/minute at 300 DPI
Continuous Runtime ~6 hours on full battery

Design Principles

Every design decision in Yantrax is driven by three constraints:

  • No Internet: Every bit of computation happens on-device. No cloud calls, no fallbacks to remote servers.
  • Real Time: Museum environments are unpredictable. The robot must perceive, decide, and act in under 200ms to feel responsive.
  • Graceful Degradation: If one sensor fails (e.g., LiDAR), the robot continues operating with reduced capability rather than shutting down.