Manvel Avetisian

Building an autonomous delivery robot from scratch in Python - why Python for robotics, and where it breaks down.

Here are interesting topics I'd like to speak about, will cut something depending on time available:
1. Hardware architecture decisions for a sub-$500 robot: Raspberry Pi, Orange Pi, stereo cameras, LiDAR, IMU, and why sensor choice determines your entire software stack.
2. Fighting Python's performance ceiling: threading for I/O-bound sensor fusion, multiprocessing for CPU-bound perception, and when to drop into C++ extensions.
3. On-device neural network inference on a $15 NPU chip: training in PyTorch, exporting to ONNX, deploying via RKNN, and the surprising finding that EfficientNet is slower than Wide ResNet at 10x the FLOPs.
4. Building a custom lightweight simulator for reinforcement learning: procedural sidewalk generation, moving pedestrians, physics, and rendering - replacing Panda3D with a purpose-built engine that runs full test suites in three minutes.
5. Training SAC policies in JAX: reward shaping, replay buffers, critic divergence, and getting from 0% to 20% success rate on autonomous sidewalk navigation.
6. The sim-to-real gap and how to close it: camera decalibration emulation, sensor noise injection, motor dead zones, perception model dropout, and why domain randomization at the model output level matters more than pixel-level randomization.
7. Deploying learned policies to real hardware: inference latency budgets, async perception pipelines, and the stop-and-wait wrapper that prevents the camera-blur death spiral.
8. War stories from field testing: battery fires, puddle submersions, snow, and the hard lesson that classical control with noisy ML perception is a recipe for disaster.