Sandeep Reddy Sabbella

CANOPIES is an H2020 European Project (Grant agreement ID: 101016906) aiming to develop a novel collaborative human-robot paradigm in the field of precision agriculture for permanent crops, where farmworkers can efficiently work with teams of robots to perform agronomic interventions, such as harvesting or pruning in table-grape vineyards.

Developed a high-fidelity 3D simulation and virtual reality (VR) evaluation framework for Embodied AI systems, enabling the systematic assessment of robotic perception, decision-making, and action capabilities in realistic, interactive environments. The framework was designed to bridge the gap between simulated training and real-world deployment by providing controlled, repeatable, and scalable testing conditions for embodied robotic agents.

The simulation environments incorporated realistic 3D scenes, physics-based interactions, dynamic objects, environmental variations, and sensor models to closely approximate real-world operating conditions. These environments enabled robots to interact with objects, navigate complex spaces, and respond to changing surroundings while being evaluated across multiple perception-action cycles. VR capabilities further supported immersive scenario creation and human-in-the-loop evaluation, allowing realistic task configurations to be designed and tested efficiently.

Developed a comprehensive data curation, preprocessing, and instruction-tuning pipeline for training multimodal and large language models (LLMs) to understand complex real-world environments from an egocentric, first-person perspective. The pipeline was designed to transform raw egocentric vision datasets into structured, high-quality training data that could support the development of embodied AI systems capable of connecting visual perception with language understanding and robotic action.

The data pipeline incorporated egocentric video, images, scene descriptions, object interactions, temporal context, and task-oriented annotations to capture how an agent perceives and interacts with its surroundings. Data curation processes focused on filtering low-quality samples, removing redundant or ambiguous observations, and selecting diverse scenarios that represented realistic human and robotic activities. Preprocessing workflows were developed to normalize visual inputs, organize temporal sequences, align observations with corresponding instructions, and generate consistent training examples.

Architected and open-sourced Robotic Simulation Benchmarking Environment (RoSmEEry), a Gazebo- and ROS-based simulation framework designed to enable systematic, reproducible, and automated benchmarking of semantic mapping algorithms for mobile robotics. The platform was developed to address the challenges of evaluating mapping systems consistently across different environments, sensor configurations, and algorithmic approaches without requiring repeated physical-world experiments.

RoSmEEry provides a configurable simulation environment in which autonomous robots can navigate realistic virtual spaces while collecting sensor data and constructing semantic representations of their surroundings. The framework integrates Gazebo for physics-based simulation and environment modeling with ROS for robot control, sensor integration, communication, and experiment orchestration. This architecture enables researchers and developers to rapidly prototype, test, and compare semantic mapping approaches within a standardized evaluation pipeline.

Designed and implemented an enterprise-scale visual similarity and multimodal vector search system for large-scale eCommerce product discovery, retrieval, and recommendation. The system was developed to enable customers and internal applications to discover visually similar products using product images, natural-language descriptions, or a combination of both, supporting efficient search across catalogs containing 1+ million SKUs.

I built e2e pipelines for generating high-dimensional image and text embeddings using modern computer vision and multimodal representation learning techniques. Product images and associated metadata were transformed into dense vector representations that captured visual attributes, product categories, styles, colors, shapes, and semantic relationships. Multimodal embeddings enabled text and image queries to be represented in a shared vector space, allowing users to perform flexible searches such as finding visually similar products, locating products that match a textual description, or combining visual and semantic intent.