Training and Deploying RL for a $500 Sidewalk Robot
Building a sub-$500 autonomous delivery robot from scratch: custom Vulkan simulator, SAC on JAX, sim-to-real transfer, and lessons from field testing.
Manvel Avetisian
Applied Robotics Researcher
I build low-cost autonomous robots and test them in real-world conditions. Before going independent, I spent 20 years in applied AI research — at Google, Yandex, AMD, Sber AI Lab, and AIRI — working on search, speech synthesis, medical AI, genomics, and chip design. At Sber and AIRI I led R&D teams of 100+ people.
Writing
Research
Experiment 01 — The Deliverator project
The Deliverator project investigates using AI to compensate for cheap hardware in autonomous delivery robots. The core question: can lightweight neural networks, running on-device, achieve reliable sidewalk navigation using only basic cameras, consumer-grade IMUs, and standard GPS — bringing total build cost under $500?
The entire stack is built from scratch by a single researcher: a custom Vulkan-based simulator, an RL training pipeline (SAC on JAX), vision-based navigation, teleoperation, and hardware. A functional sub-$500 prototype is operating in real-world outdoor conditions.
Status updates
- March 2026 — Reinforcement learning policy training in simulator (video), Autonomous driving: RL/IL + vision-based navigation deployed on device (video)
- February 2026 — Autonomous driving test runs successful (video)
Open research problems
- Reducing human intervention — improving vision-based navigation toward full autonomy
- Weather resilience — reliable operation in rain, snow, and low visibility
- Road crossing — safe road crossing with low-cost perception
- Mechanical durability — long-term reliability of sub-$500 hardware under daily use
Previous work
- AI in Healthcare — Medical image analysis and clinical decision support, deployed in large healthcare facilities. Published at MICCAI, AIME, IEEE.
- AI in Genomics — Interpreting noncoding DNA variants with deep learning.
- Speech Synthesis — 4x faster parametric TTS training; novel autoregressive neurovocoder at Yandex.
- ML for Chip Design — 10x inference speedup with minor quality loss at AMD.
- Search & IR — Ranking pipelines, knowledge panels, and ML annotation systems at Google.