Yantrax
Contact us
Stories, lessons, and insights from building an offline AI robot for cultural heritage.
When we started Yantrax, we faced a critical choice: cloud-based AI or edge computing? The answer became clear when we realized museums need robots that work without internet. Here's how we settled on the Jetson Orin Nano and what it taught us about offline inference at scale.
Building a robot's navigation stack is like building a plane—you can't debug it mid-flight. We spent three weeks tuning SLAM parameters, testing in different lighting conditions, and learning why your loop closure detector needs to be robust before you ship.
We talked to curators, conservators, and security staff. They taught us something machines often miss: a robot in a museum isn't an exhibit, it's a tool. And tools should be invisible. How we designed Yantrax to fit into heritage spaces without overshadowing them.
Running a state-of-the-art vision model on 8GB of RAM means quantization isn't optional—it's survival. We went from FP32 to INT8 while keeping 94% accuracy. This is what we learned about the dark arts of model compression.
ROS 2 is powerful, but deploying it on a robot that needs to run 6+ hours without supervision is a different beast. Real-time concerns, memory leaks, and the joys of debugging a distributed system on moving hardware.
Yantrax can greet visitors in six Indian languages. But making speech feel natural—not robotic—meant diving deep into prosody, pacing, and the subtle art of making machines sound welcoming. Our journey with text-to-speech.
We're working on more posts about our competition journey, the technical challenges we've overcome, and reflections on what it means to build robots for cultural preservation. Stay tuned.