Ilia
Larchenko.

At Agoda, I lead multiple Data Science and Machine Learning teams working across pricing, bidding, LTV attribution, experimentation, FinTech, bot detection, and LLM agents. I combine organizational leadership with hands-on technical and system-architecture work.

I also conduct independent robotics research, focusing on reinforcement learning, imitation learning, VLA models, sim2real, and real-world manipulation. Recent results include winning the NeurIPS 2025 BEHAVIOR Challenge and placing 1st in simulation and 2nd on real hardware in the ICRA 2026 LeHome Challenge. Earlier, I reached the global top-20 on Kaggle.

I publish code, technical reports, experiments, and detailed videos through GitHub, arXiv, this website, and YouTube.

02 /

Career

Full history on LinkedIn ↗

2020 — now

Director, Data Science

Agoda · Bangkok

Lead multiple DS/ML teams across pricing, bidding, LTV attribution, experimentation, FinTech, bot detection, content optimization, and LLM agents.

Directly manage ~20 people while providing technical and architectural leadership across larger cross-functional organizations. ML systems built by my teams are used by millions of customers.

2016 — 2020

Chief Innovation Officer

DOC+ · Moscow

Founded and led a 15-person DS/ML department at a health-tech startup. Built a pre-LLM medical chatbot symptom checker, electronic health record quality control and an automated doctor-dispatching system.

2011 — 2016

Earlier experience

Consulting · startups · enterprise

Five years across analytics, strategy, finance, business development and operations — from consulting to startups and corporates.

Education

2008 — 2014

M.Sc. Applied Physics & Mathematics

Moscow Institute of Physics and Technology (MIPT)

Graduated with honors. Faculty of Control and Applied Mathematics.

03 /

Research & Competitions

🥇 1st / 62 simulation🥈 2nd real-worldICRA · 2026

LeHome Challenge

Solo entry in a bimanual garment-folding challenge. In simulation, I built an end-to-end VLA + RL pipeline with asynchronous distributed training, value prediction, advantage conditioning, and inference-time optimization, finishing 1st of 62 teams. For the real-world final, I built the sim2real and DAgger-like data-collection pipeline and adapted the policy to physical hardware in about a week, finishing 2nd.

VLARLsim2real

Solution blog ↗ Paper ↗ Code ↗ Challenge ↗

🥇 1st placeNeurIPS · 2025

BEHAVIOR Challenge

Winner of the BEHAVIOR Challenge at NeurIPS 2025, organized by the Stanford Vision & Learning lab: 50 long-horizon household tasks in simulation. Our team of three trained a modified Pi0.5 VLA policy on demonstrations and released the tech report on arXiv, the code, and a detailed video.

VLAPi0.5Long-horizonImitation learning

Tech report ↗ Code ↗ Challenge site ↗

Kaggle MasterKaggle

Kaggle Competitions Master

25 competition medals, peak top-20 globally out of 150k+ users. Master tier in both Competitions and Notebooks.

Competitions MasterNotebooks Master25 medals

Profile ↗ Competitions ↗ Notebooks ↗

TOP-20

Peak global

COMPETITIONS MEDALS

Master

Competitions

Master

Notebooks

🥉 3rd placeKaggle · 2020

Abstraction & Reasoning Challenge

3rd place out of ~900 teams in the first ARC Kaggle competition — the predecessor of today's ARC-AGI benchmark. Working in a team of 2, we each developed independent solutions and the final submission merged both into a single symbolic solver: a small DSL of grid transformations plus search. No ML, AI or LLMs.

ARCKaggleProgram synthesisDSL

Repo ↗ Competition ↗

04 /

Videos

Channel · @ilialarchenko ↗

I publish detailed technical videos about robotics and AI, usually based on systems I have built myself. Topics include VLA models, reinforcement learning, imitation learning, sim2real, robot control, and competition solutions.

05 /

Selected Projects

All repos on GitHub ↗

Open sourceBehavior cloning

dot_policy

Decoder-Only Transformer behavior-cloning policy. Sometimes 100× smaller than SOTA, no diffusion, no tokenization, no RL — very simple, but beats the prior best on PushT and ALOHA Insert.

BCTransformerPushTALOHA

Code ↗

Open sourceLLM + robotics

robot_MCP

A Model Context Protocol server for the SO-ARM100 low-cost robot arm. Lets an LLM drive the arm directly through tool calls — no separate robotics policy required.

MCPSO-ARM100AgentsLLM

Code ↗ Detailed video ↗