$
cat research_statement.txt
My research focuses on building generalist robots by combining deep generative models, 3D robot learning, learning from video, and optimization. I believe this goal can only be achieved by developing novel strategies to scale the available data for robotics and finding new algorithmic and architectural elements to enhance generalization.
Across my projects, a core theme has been learning structured representations for robot manipulation — from energy-based policies and diffusion fields to equivariant flow matching and stable dynamical systems on Lie groups.
$
ls -lt ./publications/
01
▶
A comprehensive survey on deep generative models for robot learning from multimodal demonstrations,
covering diffusion models, energy-based models, normalizing flows, and their applications to
manipulation, locomotion, and dexterous control.
02
▶
A generative framework for imitation learning from observation that uses noise-conditioned
energy-based models with annealed rewards to learn robust policies without access to expert actions.
03
▶
Proposes equivariant flow matching for robot policy learning that leverages spatial symmetries
to achieve accurate and efficient manipulation policies. Won Best Paper Award at the
Structural Priors as Inductive Biases for Learning Robot Dynamics workshop at RSS 2024.
04
▶
A framework for learning dexterous piano playing from internet video demonstrations,
enabling a robotic hand to perform complex musical pieces through generalist imitation learning.
05
▶
Introduces diffusion models on SE(3) for learning smooth cost functions that jointly optimize
grasp poses and motion trajectories. Won Best Paper Award at the Geometric Representations
Workshop at ICRA 2023.
06
▶
Proposes a method for learning stable dynamical systems on Lie groups, enabling robots to
generate stable orientation-aware motions with guaranteed convergence properties.
07
▶
Introduces implicit neural representations as priors for trajectory optimization,
enabling efficient motion planning that combines learned cost landscapes with
optimization-based control.
08
▶
Proposes composable energy-based policies for reactive robot motion generation, where
multiple energy functions can be combined to achieve complex behaviors. Extended from
R:SS to IJRR.
09
▶
A comprehensive benchmark for structured policies and policy optimization methods
applied to real-world dexterous object manipulation tasks.
10
▶
Uses normalizing flows to learn deep stable stochastic dynamical systems for imitation learning,
combining the expressiveness of deep generative models with stability guarantees.
julen@research:~$