Agon Serifi

I am a Research Scientist at Google DeepMind, working on the Gemini Robotics team. My research focuses on the intersection of mind and body, creating embodied intelligences that master diverse, lifelike motions to bring their full capability and expressivity into the real world. To this end, I combine deep reinforcement learning with human motion data, developing control methods that let robots generalize and exhibit rich behaviors, from imitating to generating motion. Previously, I was an Associate Research Scientist with the Disney Research Robotics team. I received my PhD in Computer Science (2025) from the Computer Graphics Laboratory (CGL) at ETH Zurich under Prof. Markus Gross and Dr. Moritz Bächer, in collaboration with Disney Research. I also hold a Bachelor's (2019) and a Master's (2021) degree in Computer Science from ETH Zurich.

News

Loading news

Highlight

Research

Loading research papers

Kamino: GPU-based Massively Parallel Simulation of Multi-Body Systems with Challenging Topologies

Vassilios Tsounis*, Guirec Maloisel*, Christian Schumacher*, Ruben Grandia, Agon Serifi, David Müller, Chris Amevor, Tobias Widmer, Moritz Bächer

CoRL 2026 (Spotlight)

project page | paper | video | code | ArXiv

Kamino is a GPU-based physics solver for massively parallel simulation of heterogeneous, highly-coupled mechanical systems, natively supporting complex topologies with kinematic loops. Built on NVIDIA Warp and integrated into Newton, it enables batched simulation of structurally diverse robots on a single GPU for reinforcement learning.

Walk Like This: Stylized Locomotion for Physics-Based Characters through Example-Driven Diffusion

Joshua Aurand, Agon Serifi, Sammy Christen, Ruben Grandia, Espen Knoop, Marco Hutter, Moritz Bächer

CoRL 2026

project page | paper | video

We present a real-time framework for expressive legged locomotion that controls velocity and style independently, learning from large-scale unstructured motion data without style labels. A style encoder conditions an autoregressive diffusion model that turns a style reference clip and a joystick command into a kinematic reference, which a reinforcement learning controller tracks in simulation and on a physical robot.

CoCo-InEKF: State Estimation with Learned Contact Covariances in Dynamic, Contact-Rich Scenarios

Michael Baumgartner, David Müller, Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, Moritz Bächer

RSS 2026

paper | video | RSS

CoCo-InEKF is a differentiable Invariant Extended Kalman Filter that replaces binary contact states with continuous contact covariances learned end-to-end. This improves velocity estimation on a bipedal robot and enables robust, contact-rich behaviors like dancing without heuristic contact labels.

ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting

David Müller, Agon Serifi, Sammy Christen, Ruben Grandia, Espen Knoop, Moritz Bächer

SIGGRAPH 2026

paper | ArXiv

ReActor is a bilevel optimization framework that jointly adapts human reference motion to a robot's morphology while training a tracking policy with reinforcement learning. Using only sparse semantic correspondences, it eliminates manual retargeting and produces physically plausible motion even on morphologies far from human, including a quadruped.

Olaf: Bringing an Animated Character to Life in the Physical World

David Müller*, Espen Knoop*, Dario Mylonopoulos, Agon Serifi, Michael A. Hopkins, Ruben Grandia, Moritz Bächer

IEEE Robotics and Automation Letters (RA-L), presented at IROS 2026

paper (arxiv) | video | IEEE

In this work, we bring Olaf, an animated character, to life in the physical world, relying on reinforcement learning guided by animation references for control.

Robot Crash Course: Learning Soft and Stylized Falling

Pascal Strauch, David Müller, Sammy Christen, Agon Serifi, Ruben Grandia, Espen Knoop, Moritz Bächer

ICRA 2026

paper (arxiv) | video | IEEE

We learn controlled, soft falls for bipedal robots that minimize physical damage while achieving desired end poses, using a robot-agnostic reward during reinforcement learning. The method supports arbitrary end-pose specification at inference time and demonstrates robust falling in both simulation and the real world.

AMOR: Adaptive Character Control through Multi-Objective Reinforcement Learning

Lucas N. Alegre*, Agon Serifi*, Ruben Grandia, David Müller, Espen Knoop, Moritz Bächer

SIGGRAPH 2025

project page | paper | video | ACM

AMOR enables adaptive character control by training a single multi-objective reinforcement learning policy conditioned on reward weights and motion context. This allows post-training tuning of reward trade-offs, zero-shot adaptation, and sim-to-real transfer of dynamic motions across a broad range of tasks.

Autonomous Human-Robot Interaction via Operator Imitation

Sammy Christen, David Müller, Agon Serifi, Ruben Grandia, Georg Wiedebach, Michael A. Hopkins, Espen Knoop, Moritz Bächer

IROS 2025

project page | paper | video | IEEE

We autonomously control expressive human-robot interactions by training a diffusion-based imitation model on real operator data. Conditioned on human pose, it predicts continuous and discrete control inputs, producing robust, lifelike interactions and diverse moods that resemble expert operators.

Robot Motion Diffusion Model: Motion Generation for Robotic Characters

Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, Moritz Bächer

SIGGRAPH Asia 2024

project page | paper | video | ACM

We align text-conditioned kinematic motion generators with the capabilities of robotic characters by combining motion generation with physics-based control. A learned critic acts as a differentiable reward surrogate for the non-differentiable control task, aligning the generator with the character's physical capabilities.

Spline-Based Transformers

Prashanth Chandran*, Agon Serifi*, Markus Gross, Moritz Bächer

ECCV 2024 (Oral)

project page | paper | supp | video | Springer

Spline-Based Transformers are a new class of Transformers that eliminate positional encoding by embedding an input sequence as a smooth trajectory in latent space. Leveraging the continuity of sequential data, they outperform traditional transformers in efficiency and accuracy.

VMP: Versatile Motion Priors for Robustly Tracking Motion on Physical Characters

Agon Serifi, Ruben Grandia, Espen Knoop, Markus Gross, Moritz Bächer

SIGGRAPH/Eurographics Symposium on Computer Animation (SCA) 2024

project page | paper | video | ACM

VMP is a two-stage framework that transfers expressive, dynamic motion onto robotic characters by first learning a latent motion prior from human motion capture. This prior then trains a reinforcement learning tracking controller, enabling lifelike, dynamic motions such as dancing.

Transformer-Based Neural Augmentation of Robot Simulation Representations

Agon Serifi, Espen Knoop, Christian Schumacher, Naveen Kumar, Markus Gross, Moritz Bächer

IEEE Robotics and Automation Letters (RA-L), presented at IROS 2023

project page | paper | video | IEEE

This paper introduces a modular approach for neural augmentation of simulation states, improving robot simulation accuracy by targeting individual building blocks such as actuators and rigid bodies separately.

PhD Thesis

Lifelike Motions for Robotic Characters

Agon Serifi

Doctoral Thesis, ETH Zurich 2025 · Advisors: Markus Gross, Moritz Bächer

page | pdf | DOI

My dissertation on bringing lifelike, dynamic motion to robotic characters, combining physically accurate simulation, learned latent motion representations, and reinforcement-learning control to close the sim-to-real gap and transfer expressive motions such as dancing to a real humanoid.