Eadom Dessalene
I am a PhD Candidate in Computer Science at the University of Maryland, College Park, where I work with the Perception and Robotics Group and am advised by Yiannis Aloimonos.
My research studies how robots can learn from human videos and how visual representations learned from passive observation can be improved through embodied physical interaction. Broadly, I am interested in embodied action understanding, robot learning, egocentric vision, and multimodal perception.
PhD Candidate
Perception and Robotics Group
Department of Computer Science
University of Maryland
Perception and Robotics Group
Department of Computer Science
University of Maryland
News & Talks
Jan 2026 Invited talk: Embodied Action Understanding — NASA JPL Vision Seminar, NASA Jet Propulsion Laboratory
Dec 2025 Invited talk: Generative Models of Action — NVIDIA Research Radar, NVIDIA
Nov 2024 Invited talk: Understanding Actions from Video — NYC Computer Vision Day
Sep 2024 Invited talk: Learning the Organization of Action — University of Maryland, Baltimore County
Jun 2024 Invited talk: Learning the Organization of Action — Telluride Neuromorphic Workshop
May 2023 Invited talk: Understanding Actions from Video — CoRL Cognitive Science Workshop
Selected Publications
FEEL (Force-Enhanced Egocentric Learning): A Dataset for Physical Action Understanding
ArXiv 2026
FEEL is the first large-scale egocentric dataset pairing video with synchronized force measurements from custom piezoresistive gloves. The dataset contains approximately 3 million force-synchronized frames of natural, unscripted kitchen manipulation and introduces force as a physically grounded supervisory signal for contact understanding and action representation learning.
EmbodiSwap for Zero-Shot Robot Imitation Learning
arXiv, 2025
We introduce EmbodiSwap, a method for producing photorealistic synthetic robot overlays on human video. The approach helps bridge the embodiment gap between in-the-wild egocentric human videos and a target robot embodiment, enabling zero-shot imitation learning for robot manipulation.
Context in Human Action through Motion Complementarity
WACV 2024
We propose a learning framework in which context is modeled as the complement of motion. Physical movement is represented through Therbligs, while context is captured using a contrastive mutual-information-based objective.
LEAP: LLM-Generation of Egocentric Action Programs
arXiv 2023
LEAP generates video-grounded action programs composed of sub-actions, conditions, and control flows. It uses large language models to combine program knowledge with multimodal evidence from egocentric videos.
Therbligs in Action: Video Understanding through Motion Primitives
CVPR 2023
We introduce a compositional and hierarchical framework for action understanding based on Therbligs as motion primitives, along with differentiable rule-based reasoning for logical consistency.
Forecasting Action through Contact Representations from First-Person Video
PAMI 2021
We develop contact-centered representations and models for first-person video, motivated by the role of hand-object contact in the structure and anticipation of human action.
Using Geometric Features to Represent Near-Contact Behavior in Robotic Grasping
ICRA 2019
We define hand-object geometric feature representations for robotic grasping at the near-contact stage, designed to be robust to noise, morphology differences, and suitable for direct machine learning use.
Service
Reviewer: ICRA, ICLR, CVPR, PAMI, WACV
Competitions
EPIC-Kitchens Action Recognition Challenge 2024 — 4th Place
Alexa Prize SimBot Challenge — Team Lead, University of Maryland (Qualified for Semi-Finals)
EPIC-Kitchens Action Anticipation Challenge 2020 — 1st Place
Amazon Robotics Challenge — Qualified for Finals