<link>https://relo02.github.io/</link></image><item><title>TalOS Robotics AI — Physical AI for Humanoids</h1> <article> <h1>TalOS Robotics AI — Physical AI for Humanoids</h1> <p>Thu, 01 Jan 2026 00:00:00 +0000</p> <p><strong>Co-Founder & CTO.</strong> TalOS Robotics AI is an early-stage startup enhancing humanoid and general legged-robotics technology — building software that lets operators <strong>train, deploy, and interact</strong> with robots through human feedback.</p> <h2 id="what-im-building">What I’m building</h2> <ul> <li><strong>VLA skill libraries</strong> — Vision-Language-Action policies so a single robot can learn many skills and <em>choose the right one</em> for each industrial task, generalizing beyond a single demonstration.</li> <li><strong>An MCP (Model Context Protocol) app</strong> that wraps training, deployment, and human-in-the-loop feedback as agent-callable tools, drastically simplifying the operator workflow and lowering the skill barrier for non-experts.</li> <li><strong>The full robotics/ML stack</strong> — humanoid simulation (MuJoCo, Isaac Sim), ROS 2 integration, LLM/agent-driven control, and the dashboard tooling around it.</li> </ul> <p><code>Python</code> · <code>PyTorch</code> · <code>ROS 2</code> · <code>MuJoCo</code> · <code>Isaac Sim</code> · <code>MCP</code> · <code>LLM agents</code></p> </article> <article> <h1>Go2 Autonomous Navigation — Learned Nonlinear MPC</h1> <p>Thu, 01 Jan 2026 00:00:00 +0000</p> <p>A data-efficient framework that <strong>learns the cost weights of a nonlinear MPC</strong> navigation stack using <strong>Bayesian Optimization</strong>, treating closed-loop performance as a black-box objective and deploying the result on a <strong>Unitree Go2</strong> quadruped.</p> <ul> <li>📄 Paper: — <em>IEEE ICRA 2026 Workshop</em> (1st author)</li> <li>💻 Code: </li> </ul> <blockquote class="border-l-4 border-neutral-300 dark:border-neutral-600 pl-4 italic text-neutral-600 dark:text-neutral-400 my-6"> <p>Add a demo clip at <code>static/videos/go2-navigation.mp4</code> to play it in the homepage showcase.</p> </blockquote> </article> <article> <h1>MoonBot — YC Robotics Hackathon (w/ Innate)</h1> <p>Fri, 01 May 2026 00:00:00 +0000</p> <p>Built <strong>MoonBot</strong> (with R. Feingold, G. Voss, L. Knak, N. Rodriguez) at the <strong>Y Combinator Robotics Hackathon</strong> — sponsored by NASA, DeepMind, Scale AI, Nebius, ElevenLabs, Dryft, and Iterate.</p> <p>An autonomous robot for <strong>inspection / intervention in space-like environments</strong> under a fully embedded, <strong>edge-first architecture</strong> (no cloud, no external compute).</p> <ul> <li><strong>Perception:</strong> AprilTag-based target localization from camera feeds.</li> <li><strong>Interaction:</strong> once a target is reached, an <strong>ACT (Action Chunking Transformer)</strong> policy trained by imitation learning maps perception directly to action.</li> <li><strong>Navigation/control</strong> designed to be robust to partial observability.</li> </ul> <p><code>Python</code> · <code>PyTorch</code> · <code>OpenCV</code> · <code>ROS 2</code></p> </article> <article> <h1>GPS-Denied Drone Autonomy — EKF, ORB-SLAM3 & MPC</h1> <p>Sun, 01 Sep 2024 00:00:00 +0000</p> <p>Co-developing an autonomous drone platform for <strong>GPS-denied environments</strong> with the AEA student team at Politecnico di Milano.</p> <ul> <li><strong>State estimation:</strong> an <strong>EKF</strong> fusing IMU, GPS, barometer, magnetometer, and <strong>stereo visual odometry (ORB-SLAM3)</strong> pose estimates, deployed on an <strong>NVIDIA Jetson Orin Nano</strong>.</li> <li><strong>Planning & control:</strong> an <strong>MPC-based path planner (CasADi)</strong>, validated in <strong>Gazebo Harmonic</strong> with ROS 2 and deployed on <strong>Pixhawk/PX4</strong>.</li> <li>Leading the bring-up from simulation to real-world flight.</li> </ul> <p><code>C++</code> · <code>CUDA</code> · <code>Python</code> · <code>ROS 2</code> · <code>CasADi</code> · <code>PX4</code></p> </article> <article> <h1>Histopathology Image Classification — 4th / 170</h1> <p>Mon, 01 Dec 2025 00:00:00 +0000</p> <p>Image classification of molecular structures in tissue images for the <strong>Artificial Neural Networks & Deep Learning</strong> course at Politecnico di Milano — <strong>ranked 4th of 170 teams</strong> on a hidden test set.</p> <ul> <li><strong>Backbone:</strong> transfer learning with a <strong>ResNet-50</strong> pre-trained via <strong>RetCCL</strong> for histopathology-specific feature extraction.</li> <li><strong>Pipeline:</strong> reproducible <strong>PyTorch Lightning</strong> training with <strong>WeightedRandomSampler</strong> for class imbalance, <strong>AdamW</strong>, label smoothing, cosine annealing, and rotation/flip/color-jitter augmentation.</li> </ul> <p><code>Python</code> · <code>PyTorch</code> · <code>PyTorch Lightning</code> · <code>CNNs</code></p> </article> <article> <h1>Bayesian Optimization for Learning Nonlinear MPC in Autonomous Agent Navigation</h1> <p>Thu, 01 Jan 2026 00:00:00 +0000</p> <p>First-author work (with G. Voss, G. Beltrami, F. Dorati, and T. F. Banfi) presented at the <strong>IEEE ICRA 2026 Workshop</strong>. We learn the weights of a nonlinear MPC navigation controller via <strong>Bayesian Optimization</strong>, treating closed-loop performance as a black-box objective and deploying the result on a <strong>Unitree Go2</strong> quadruped.</p> <ul> <li>📄 </li> <li>💻 </li> </ul> </article> <article> <h1>Drone trajectory planning</h1> <p>Mon, 03 Nov 2025 00:00:00 +0000</p> <h2 id="problem">Problem</h2> <p>Autonomous navigation in GPS-denied or dynamic environments requires robust local path planning that can react to obstacles in real-time while optimizing trajectory efficiency.</p> <h2 id="approach">Approach</h2> <ul> <li>Implemented <strong>Model Predictive Control (MPC)</strong> for local trajectory optimization</li> <li>Integrated with ROS 2 navigation stack with gazebo simulation and PX4 low levels controllers</li> <li>Tested on a Quadcopter</li> <li>Combined with SLAM</li> </ul> <h2 id="technologies">Technologies</h2> <ul> <li><strong>Frameworks:</strong> ROS 2, MPC library CasADi</li> <li><strong>Sensors:</strong> LiDAR, IMU</li> <li><strong>Hardware:</strong> Quadcopter with pixhawk flight controller and Jetson Orin Nano used for higher computations</li> </ul> </article> <article> <h1>Time Series Classification for Pain Detection</h1> <p>Mon, 15 Jan 2024 00:00:00 +0000</p> <h2 id="problem">Problem</h2> <p>Autonomous pain classification in time-series data requires robust temporal modeling that can capture both short-term fluctuations and long-term dependencies from 30 joint measurements and 4 pain survey responses.</p> <h2 id="architecture">Architecture</h2> <h3 id="network-overview">Network Overview</h3> <div class="mermaid">flowchart TD subgraph Input I1[Time Series Input] I2[30 Joint Measurements] I3[4 Pain Surveys] I4[3 Prosthetic Features] end subgraph Preprocessing P1[Sliding Window] P2[Window Size: 10] P3[Stride: 2] end subgraph FeatureSplit F1[Continuous Features] F2[Categorical Features] end subgraph Encoders E1[1D Conv Layer] E2[BatchNorm + ReLU] E3[Dropout 0.6] E4[Embedding Layers] end subgraph RNN R1[GRU Layer 1] R2[GRU Layer 2] R3[GRU Layer 3] R4[Hidden: 64] R5[Dropout: 0.4] end subgraph Attention A1[Attention Weights] A2[Context Vector] end subgraph Output O1[FC Layer] O2[Softmax] O3[3 Classes] end I1 --> P1 I2 --> P1 I3 --> P1 I4 --> P1 P1 --> P2 P2 --> P3 P3 --> FeatureSplit F1 --> E1 E1 --> E2 E2 --> E3 F2 --> E4 E3 --> R1 E4 --> R1 R1 --> R2 R2 --> R3 R3 --> R4 R4 --> R5 R5 --> A1 A1 --> A2 A2 --> O1 O1 --> O2 O2 --> O3 </div> </article> <article> <h1>Experience</h1> <p>Tue, 24 Oct 2023 00:00:00 +0000</p> </article> </main></body></html>