PhD researcher at the University of Genoa working on autonomous robot navigation, deep reinforcement learning, and Sim2Real transfer. My work spans the full pipeline — from training distributed RL policies in Isaac Lab to deploying and validating them on physical UGVs via ROS2 Nav2 integration.
Prior to robotics, I completed an M.Tech in Aerospace Engineering (Dynamics & Control) at IIT Bombay, where I worked on multispectral imaging systems for remote sensing. That background now feeds directly into terrain classification and oil-leakage detection research using hyperspectral cameras on ground robots.
I teach PDDL-based AI planning as a Teaching Assistant at UniGe, and I take on selected freelance contracts in robotics software and RL engineering.
Open To: Remote robotics/AI research collaborations · RL engineering contracts · Workshop and curriculum design partnerships
| Domain | Proficiency | Details |
|---|---|---|
| Deep Reinforcement Learning | ███████████░ Expert | PPO, SAC, TD3 — custom reward shaping, curriculum scheduling |
| Sim2Real Transfer | ███████████░ Expert | Domain randomization, physics fidelity tuning, Isaac Lab → ROS2 |
| Robot Navigation | ██████████░░ Advanced | DRL local planners, Nav2 integration, costmap tuning |
| Computer Vision | █████████░░░ Advanced | ResNet, dual-stream CNNs, multispectral classification |
| Multispectral Imaging | █████████░░░ Advanced | MicaSense RedEdge-P, oil-on-soil detection, NIR/RE band fusion |
| Classical Planning | ███████░░░░░ Proficient | PDDL+, ENHSP solver, temporal planning |
| Motion Planning | ███████░░░░░ Proficient | Nav2 planners, A*, trajectory optimization |
🤖 Husky A200 DRL Local Planner — Isaac Lab → ROS2 Nav2
A production-grade deep RL local planner for the Clearpath Husky A200 UGV, trained in NVIDIA Isaac Lab using the DirectRLEnv interface and deployed as a drop-in Nav2 controller plugin via a custom ROS2 bridge.
| Attribute | Details |
|---|---|
| Stack | Isaac Lab 5.1, SKRL (PPO), PyTorch, ROS2 Humble, Nav2 |
| Training Scale | 4096 parallel envs, GPU-vectorized rollouts |
| Policy | MLP actor-critic, 480-dim observation space, SE(2) action |
| Sim2Real Gap | Domain-randomized friction, wheel slip, sensor latency |
| Deployment | Custom ControllerServer plugin, /cmd_vel direct output |
| Status | Active — sim validation complete, hardware transfer in progress |
The environment encodes goal-relative pose, LiDAR sector aggregates, and proprioceptive velocities into a compact observation, with shaped rewards for goal progress, heading alignment, and collision avoidance. Physics randomization covers floor friction (0.3–1.2), actuator damping, and IMU noise — critical for narrowing the reality gap on the physical Husky platform.
🛢️ DRIFT — Dual-Stream ResNet for Oil-on-Soil Classification
DRIFT (Dual-stream ResNet Image Fusion Technique) is a multispectral classification architecture targeting oil-leakage detection on soil surfaces using MicaSense RedEdge-P imagery. Submitted to IEEE AIM.
| Attribute | Details |
|---|---|
| Stack | PyTorch, ResNet-50 (dual-stream), scikit-learn |
| Sensor | MicaSense RedEdge-P (5-band: Blue, Green, Red, Red Edge, NIR) |
| Task | Binary oil-on-soil classification from multispectral orthomosaics |
| Architecture | Parallel CNN streams fused at penultimate layer |
| Venue | IEEE AIM (International Conference on Advanced Intelligent Mechatronics) |
| Status | Under review |
The dual-stream design processes RGB and NIR/RE band subsets independently before late fusion, exploiting the spectral contrast of hydrocarbon contamination in the near-infrared range — a signature invisible to standard RGB cameras.
🗺️ MiR 250 ROS2 Nav2 Stack with Graph-Based Navigation
Full ROS2 Humble navigation stack for the Mobile Industrial Robots MiR 250 AMR, featuring a topological graph-based mission layer on top of Nav2's metric planning pipeline.
| Attribute | Details |
|---|---|
| Stack | ROS2 Humble, Nav2, Python |
| Platform | MiR 250 AMR |
| Navigation | Global A* + DWB local planner + topological graph layer |
| Localization | AMCL with tuned particle filter parameters |
| Use Case | Autonomous warehouse-style waypoint missions |
| Portfolio Use | Freelance contract deliverable |
PhD Researcher — Autonomous Robotics · University of Genoa · Research at the intersection of deep reinforcement learning and autonomous mobile robot navigation. Develop Sim2Real pipelines using NVIDIA Isaac Lab and validate policies on physical ground robots via ROS2.
- Designed and trained DRL local planners (PPO/SKRL) for UGV navigation in Isaac Lab with 4096+ parallel environments
- Architected custom
DirectRLEnvwrappers and Nav2 controller plugins for hardware deployment - Led multispectral oil-leakage detection research using MicaSense RedEdge-P imagery and dual-stream CNNs
- First-author IEEE AIM submission (DRIFT architecture); co-author on HAN review paper (Robotics & Autonomous Systems)
Isaac Lab ROS2 PPO SKRL PyTorch Sim2Real Nav2 Multispectral CV
Teaching Assistant — AI Planning (PDDL) · University of Genoa ·
Deliver lab sessions and tutorials for the graduate AI Planning course. Design exercises in PDDL+, debug student models with ENHSP and other temporal planners, and maintain course toolchain infrastructure.
PDDL+ ENHSP Temporal Planning AI Education
M.Tech Researcher — Aerospace Dynamics & Control · IIT Bombay · 2022 – 2024
Multispectral imaging pipeline development for airborne remote sensing. Collaborated with Hmeandra Arya on dataset construction and classification architectures for vegetation and soil analysis.
Multispectral Imaging Python Remote Sensing Dynamics & Control
| Recognition | Details |
|---|---|
| 🏛️ IEEE AIM Publication | First-author submission — DRIFT dual-stream oil detection architecture |
| 📖 RAS Journal Co-Author | Human-Aware Navigation (HAN) review — Robotics & Autonomous Systems |
| 🎓 IIT Bombay M.Tech | Aerospace Engineering — Dynamics & Control |
| 🤖 Production RL Deployment | DRL planner contracted for physical UGV hardware deployment |
| 🇮🇹 UniGe PhD Scholarship | Fully-funded doctoral position, University of Genoa |
research:
building: "Sim2Real DRL local planner — Husky A200 hardware validation"
writing: "IEEE AIM DRIFT paper (under review) + HAN journal revision"
exploring: "Terrain-adaptive navigation policies + legged robot locomotion"
tools: ["Isaac Lab 5.1", "SKRL", "ROS2 Humble", "NVIDIA Warp"]
open_to:
- Research collaborations (navigation, perception, Sim2Real)
- Workshop design partnerships (robotics, AI, RL curriculum)
- Co-authorship on relevant robotics / CV papers
From spacecraft dynamics to autonomous ground robots — the physics of control doesn't care which domain you're in.