Instance segmentation on the Kaggle Fruits dataset using a Mask R-CNN, I coded the backbone from scratch. Backbone: ResNet-13 · Dataset: Kaggle Fruits · Task: apple/banana/orange instance masks
Click to view full image.Notes on the model
Built step by step: Mask R-CNN trained on the Kaggle Fruits set (apple, banana, orange). I resize everything to 200×200 for speed.
Backbones I tried: a small CNN, a ResNet-style net, ResNet-13, and a ResNet-13 with Hadamard residuals. You can switch with CustomBackbone_list_n; backbone.out_channels is set accordingly (e.g., 2048).
Proposals & pooling: custom AnchorGenerator and MultiScaleRoIAlign, with tweakable IoU/NMS thresholds for both the RPN and ROI heads.
Data pipeline: loads preprocessed .pth files (images, bbox, labels, optional masks). Works with rectangle or polygon masks and can add Gaussian noise during training.
Training setup: Adam with L2 weight decay, TensorBoard charts (train/val loss, IoU, label accuracy), optional early-stop on val accuracy, rolling epoch checkpoints, plus a separate “best model” save.
Resume & evaluate: can reload from a checkpoint or full model. The eval script runs NMS and shows ground truth vs. predictions side-by-side with mask overlays and confidence filtering.
Demo Video – Reinforcement Learning (PPO) for Bipedal Walker
Demonstration of training and evaluating an AI agent to learn walking behavior
using Proximal Policy Optimization (PPO) in a physics-based simulation.
Notes on what is demonstrated
Policy Learning: The agent learns to walk through trial-and-error interactions with the environment.
Continuous Control: Actions are smooth and continuous, suitable for realistic robotic motion.
Reward-Based Training: The agent is guided by rewards that encourage stable and forward movement.
Deterministic Evaluation: Final performance is evaluated using mean actions for consistent results.
Checkpointing: The best-performing models are automatically saved during training.
Reproducibility: Training and evaluation are designed to be repeatable and verifiable.
Demo Video – Reinforcement Learning (PPO) for Car Racing
Demonstration of training and evaluating an autonomous driving agent using
Proximal Policy Optimization (PPO) in the CarRacing-v3 simulation environment.
The agent learns directly from raw visual input to control steering, throttle,
and braking.
Notes on what is demonstrated
Visual Policy Learning: The agent learns driving behavior from RGB images without access to privileged state variables.
Continuous Control: Steering, throttle, and braking actions are continuous and smoothly applied.
Reward-Guided Driving: Training rewards encourage progress along the track, stability, and lap completion.
Deterministic Evaluation: Final demonstrations use deterministic policy execution for consistent behavior.
Automatic Checkpointing: The best-performing models are saved during training based on evaluation performance.
Reproducibility: Training and evaluation are structured to produce repeatable and verifiable results.
Demo Video - Laser Powder Bed Fusion (LPBF)
Short demo of my Laser Powder Bed Fusion (LPBF) setup using a custom Python/pygame GUI to coordinate XYZ linear stages and a galvo scanner in real time.
Demo Video - ABB Robot Absolute & Relative Movements
Demonstration of ABB robot executing various motion modes using RAPID instructions.
Notes on performed movements
Absolute Axis Movement: Direct movement of individual robot axes to specified absolute positions.
Absolute Linear Movement: Straight-line motion between points in Cartesian space.
Absolute Circular Movement: Circular interpolation movement between points along a defined arc.
Relative Axis Movement: Movement relative to the robot’s current joint angles.
Relative Linear Movement: Translation by a specified offset from the current Cartesian position.
Absolute Linear Movement synchronized with positioner 2-axis movement: Coordinated robot and positioner motion for complex toolpath execution.
Defense Manufacturing Expo — UTRGV (2022)
Lead-through teaching demo: I moved the robot by hand while the system recorded waypoints to a file, then replayed them automatically (including pick-and-place). Entire workflow in Python.
URDF Robot Demo
A live 3D viewer of one of my robot model (URDF). You can rotate, zoom, and move the joints.