Lead Researcher at T-Tech, Moscow
I work on embodied AI: world models and vision-language-action policies for robots and virtual worlds.
I chase the same questions in two domains at once, robots and virtual worlds. The through-line is that hours of video on the internet outnumber recorded hours of robot action by a factor of thousands, and teleoperation scales badly — so I look for recipes that grow a model from cheap video, from its own rollouts under an automatic judge, and from generated tasks that check themselves, rather than from new demonstrations alone. In practice that means world and action models that learn how a world answers an action, and vision-language-action policies that turn a camera image and a plain-language instruction into behaviour. Earlier my research was in 3D computer vision and neural rendering.
I hold a Ph.D. in Computer Science, which I earned at Skoltech within Evgeny Burnaev's Applied AI Center. I received a Master's Degree in Data Science from Skoltech in 2020. Prior to that, I received my Bachelor's in Applied Mathematics and Physics from the Moscow Institute of Physics and Technology (MIPT) in 2018.
ICML Workshop (DEMO), 2026
We train a single JEPA world model whose objective pairs a Brownian-bridge interpolant on the state axis with flow matching on the action axis, so one checkpoint serves latent planning, behaviour-cloning action sampling and inverse dynamics without retraining.
2026
ICML Workshop (RLxF), 2026
We train a VLM to rank shuffled frames by task progress, then reward a policy by the rank correlation between predicted progress and real time — a bounded, scale-free signal that needs no environment reward and transfers across tasks.
ICLR Workshop, 2026
A decoder-free model-based RL agent that predicts next-step encoder embeddings from latent state sequences, matching or exceeding DreamerV3 on the DeepMind Control Suite and improving substantially on DMLab memory and spatial reasoning tasks.
2026
AAMAS, 2026
A lightweight RL recipe for vision-language agents that applies PPO at the token level for actions while learning a step-level value function. Training in one cheap synthetic environment produces policies that transfer beyond the training simulator.
2025
IROS (Oral), 2025
We present a novel approach that integrates dense keypoint descriptors into 3D Gaussian Splatting to enhance visual localization, achieving state-of-the-art performance on popular indoor and outdoor benchmarks.
2024
Preprint, 2024
We propose a novel framework that leverages 3D Gaussian Splatting to effectively remove transient objects from input videos, enabling accurate and stable 3D scene reconstruction.
CVPR, 2022
We take the original NPBG pipeline and make it work without per-scene optimization.
2022
SIGGRAPH, 2022
Differently from existing data-driven methods for predicting sharp geometric features in sampled 3D shapes, which reduce this problem to feature classification, we propose to regress a scalar field representing the distance from point samples to the closest feature line on local patches.
IEEE Access, 2023
We present an approach for the reconstruction of textured 3D meshes of human heads from one or few views.
2023
CVPR, 2023
A new multi-sensor dataset for 3D surface reconstruction that includes registered RGB and depth data from sensors of different resolutions and modalities under a large number of lighting conditions.
2021
WACV, 2021
We target the problem of redesigning the DensePose R-CNN model's architecture so that the final network retains most of its accuracy but becomes more light-weight and fast.
VISIGRAPP, 2021
We predict future video frames in latent space in an autoregressive manner.
Patches merged upstream:
--num-frames terminate a headless GL run, and documenting the pyglet
requirement behind it.
PRs
Open and merged, everywhere else: every public pull request I've sent upstream.