CV
Curriculum vitae of Luca Nunziante.
Contact Information
| Name | Luca Nunziante |
| Professional Title | Robot Learning Researcher |
| luca.nunziante1999@gmail.com |
Professional Summary
Robot learning researcher working on Vision-Language-Action (VLA) models and imitation/reinforcement learning for manipulation, with hands-on experience taking policies from simulation to real-world deployment on physical robots. Comfortable across the full stack: low-level control, teleoperation and data collection, model training, and deployment infrastructure.
Experience
-
2024 - present Tokyo, Japan
Robot Learning Researcher
Araya Inc.
- Develop imitation learning and reinforcement learning pipelines for robot manipulation built on VLA models.
- Built end-to-end training infrastructure for IL and RL on custom datasets, improving manipulation success rate by iterating and fixing edge cases encountered during deployment.
- Implemented low-level controllers (operational-space, CBF-based) and teleoperation interfaces enabling safe data collection and reliable real-world policy execution.
- Built custom MuJoCo simulation environments and the deployment infrastructure that transfers trained policies to physical hardware.
-
2023 - 2024 Sendai, Japan
Visiting Research Student
Tohoku University
- Developed compliant control methods for robotic assembly tasks and perception systems for lunar applications.
- Built a dual-arm solar-panel assembly pipeline (MSc thesis project) that led to a peer-reviewed publication (SII 2025).
Education
-
2021 - 2024 Rome, Italy
MSc
Sapienza University of Rome
Artificial Intelligence and Robotics
- Thesis: Autonomous Assembly of Solar Panels by Dual Robot Arms in the Remote Space (Supervisors: Prof. A. De Luca, Prof. K. Yoshida).
-
2018 - 2021 Caserta, Italy
BSc
University of Campania Luigi Vanvitelli
Electronic and Computer Engineering
- Thesis: Object detection with Neural Networks for robotics applications: training methods compared (Supervisor: Prof. C. Natale).
Projects
-
VLA Latent Space Investigation
Following an existing paper, I set up a failure detection system for the VLA based on the VLA internal features. Then, I wondered if those features are only informative or can also be controlled in some way: I investigated some ways to intervene in the latent space to reduce the failure cases.
-
Advantage Conditioned VLA Learning
After training a VLA, I collected rollouts and, following an existing work, I passed the frames to a pretrained VLM to predict value labels in a zero-shot manner. Then, I converted those values in advantages and retrained the VLM conditioning on those advantages. However, after inspection, the quality of the produced annotation was too poor and noisy to yield performance improvements.