Manvel Avetisian
Applied Robotics Researcher
I’m an applied researcher with 20 years of experience as a research engineer and technical leader, including Google, Yandex, Sber AI Lab, AMD, and AIRI. My work has ranged from search infrastructure and quality, speech synthesis, and medical AI to genomics and chip design. Most recently, I led research teams of 100+ people at Sber and AIRI focused on applied AI research and deployment.
Currently I’m doing independent applied research on low-cost autonomous robotics, tested in real-world conditions.
Research
Experiment 01 — The Deliverator project
Existing delivery robots cost $2,000–5,000 per unit and rely on expensive sensor suites — LiDAR, high-resolution cameras, precision IMUs — totaling $1,000+ in sensing hardware alone. This makes autonomous delivery economically viable only in dense, high-income urban areas.
The Deliverator project investigates a different approach: using AI to compensate for cheap hardware. The core research question is whether modern lightweight neural networks, running on-device, can achieve reliable sidewalk navigation using only basic cameras, consumer-grade IMUs, and standard GPS — bringing total build cost under $500.
Prototype status
A functional prototype is built and operating in real-world conditions:
- Sub-$500 build cost achieved with off-the-shelf components
- On-device neural network inference on a low-power SoC
- Autonomous navigation with obstacle avoidance and path planning
- Remote monitoring and teleoperation via internet connection
- First tests outdoors in winter conditions (snow, ice)
Open research problems
- Reducing human intervention. The central AI challenge: improving vision-based navigation to bring down teleoperation interventions, then toward full autonomy. Each reduction directly improves operational viability.
- Weather resilience. Enabling reliable operation in rain, snow, and low-visibility conditions through robust perception models and weatherproofed hardware design.
- Road crossing. The prototype currently operates sidewalk-only. Safe road crossing requires a different class of perception and decision-making — this is a hard open problem for low-cost configurations.
- Mechanical durability. Suspension, chassis weatherproofing, and long-term reliability of sub-$500 hardware under daily operational stress. The design must remain simple enough to be built and maintained by technicians with basic skills.
Previous work
Over the past 15 years, my research has spanned several applied AI domains — from healthcare and genomics to speech and search systems. Recent focus has shifted to robotics:
- AI in Healthcare — Medical image analysis (stroke detection, COVID-19, tumors) and clinical decision support systems, deployed in large healthcare facilities. Published at MICCAI, AIME, and in IEEE journals.
- AI in Genomics — Interpreting noncoding DNA variants with deep learning.
- Speech Synthesis — Improved parametric TTS training 4x and developed a novel autoregressive neurovocoder at Yandex.
- ML for Chip Design — 10x model inference speedup with minor quality loss at AMD.
- Search & Information Retrieval — Ranking pipelines, knowledge panels, and ML annotation systems at Google.