Projects

How far can 48 phone videos of my own hands take a robot? SmolVLA learns to water plants in simulation: 96% grasp, 50% full task.

Language-conditioned queries score the native X-VLA image-token grid to expose which regions matter for the current manipulation step.

Recover expert perceptual structure from X-VLA representations such as DINOv2-like tokens or CeDiRNet-aligned maps to support action guidance and representation alignment.

PIU protects a chosen identity in ID-conditioned diffusion models by re-teaching what its identity conditioning means, so the model stops generating that person while image quality stays intact.

Case study: trailer loading with multi-SKU shipments.

Two-stream encode–attend–fuse architecture that ingests global operation features and per-item (sub-operation) features, and predicts both a baseline duration and its uncertainty σ with a heteroskedastic Gaussian head.

Implementation of Deep Deterministic Policy Gradient (DDPG) with Hindsight Experience Replay (HER) for robotic manipulation task using xArm6.

• 1D LiDAR for object detection.
• Real-time encoder odometry fused with landmark detections for localization using an Extended Kalman Filter (EKF).
• Image-Based Visual Servoing to reach ArUco blocks.
• Smooth differential-drive control law for motion and reactive collision avoidance using Bug0 and Bug2 algorithms.

Gazebo simulation of a differential drive robot, fusing wheel-encoder odometry with LIDAR landmark (corner) detections using an Extended Kalman Filter (EKF).

The EKF corrects pose estimation with measurement updates from detected corners for continuous, drift-reduced localisation.

Implementation of a control strategy mapping image-point errors to differential-drive commands, following Siradjuddin et al.

Implementation of Samsung’s SceneGrasp model for multi‑object 3D reconstruction, 6DoF pose, and grasp prediction in an UR5 robot and external RGB‑D camera setup .

Design of the ROS1 system considering robot kinematics and multi-frame of reference translations.

Implementation of a self-trained Deep Deterministic Policy Gradient (DDPG) model for xARM6 tilt-stable navigation and manipulation tasks using ROS.

Implementation of visual servoing techniques for the xArm6 robotic arm, enabling precise control based on external camera's visual feedback.

Development of an self-driving vehicle prototype using ROS1, YOLOv5, and Nvidia Jetson Nano. Autonomous navigation system prototype using CV-based traffic signals detection and line following with adaptive linear control.

A prototype for an intelligent air pressure control system for tires in agricultural equipment using linear (PID) control strategies.

Developed using an ARM cortex processor in a dual core STM32 board. Utilized CAN, and serial protocols for sensor communication and a NodeMCU device for internet connectivity and online data visualization.