Papers · Talks · Posters · Projects
Publications, Talks & Projects
Peer-reviewed work, research talks and posters, and selected projects in robot manipulation, planning, learning, and verification.
Publications
Peer-reviewed publications are coming soon.
Talks & Posters
Manipulation Planning with Convex Decompositions of C-free and NLP-Sampling Talk
- Presented at an IPAB Workshop, an informal seminar series for IPAB staff and students to share ongoing work with the institute's research community.
- Introduced my manipulation-planning work: decomposing collision-free space into convex regions, planning with Graphs of Convex Sets, and extending the representation to grasp-feasible configurations with nonlinear-programming sampling.
Poster and project pitch on long-horizon manipulation planning Poster & Pitch
- Presented a poster and a short project pitch on my PhD research.
Research Projects
Manipulation Planning with Convex Decompositions
Reachability representations for long-horizon manipulation
- Extending the planner beyond collision-free motion toward grasp, placement, kinematic, and task-constrained feasible sets.
- Combining Drake inverse kinematics and SceneGraph collision checking with nonlinear hit-and-run sampling to generate diverse full-arm feasible configurations.
- Fitting mode- and submode-specific convex regions, studying transition sets between them, and planning over the resulting Graph of Convex Sets.
Planning over convex decompositions of free configuration space
- Developed a manipulation planner using IRIS variants, Graphs of Convex Sets, and Mixed-Integer Quadratic Programming.
- Implemented and visualised the planner using the Drake Robotics Toolbox.
- Awarded a First-Class mark.
A tangent configuration-space approach
- Formalised tangent parameterisations for robot configuration spaces and applied C-IRIS to manipulation planning.
- Demonstrated the scalability of C-IRIS to robots with universal joints; awarded a First-Class mark.
Neural Network Control and Verification for Tethered Drone Delivery
- Built the full neural-network training pipeline for a tethered drone-delivery controller.
- Applied Lipschitz-oriented robustness training to the learned controller.
- Specified vehicle verification properties and verified them using Marabou.
- Joint work with Calum Arnott and Ruchira Ray.
Monte Carlo Tree Search for Robotics Manipulation (Voluntary Project)
- Explored the use of MCTS for manipulation in known benchmarks such as Gymnasium or Metaworld.
- Focused on understanding the integration of Reinforcement Learning (RL) concepts, Monte-Carlo simulations, and tree search algorithms with discretised manipulation scenarios.
- Implemented AlphaZero from scratch, achieving policy generation for the selected benchmarks.
- Supervision: Dr. Nicolas Mansard (LAAS-CNRS) and Prof. Ludovic Righetti (NYU).
Object Segmentation Techniques: UNet, Autoencoder, CLIP, and Point-Based Segmentation
- Collaborated in a team of two to implement and analyze different object segmentation techniques.
- Implemented UNet, Autoencoder, CLIP, and Point-Based Segmentation techniques using PyTorch.
- Utilized Nvidia GPU servers to accelerate model training.
- Conducted a comparative analysis of the performance of each technique on the Oxford III Pet Dataset.
- Achieved a Dice score of over 80% for the CLIP model.
IoT Systems for Live Activity Classification using CNNs
- Used Tensorflow and Keras to train different CNN models for activity and respiratory condition classification.
- Utilised Android Studio to develop an app for data collection and live classification.
- Achieved an accuracy of over 90% for activity detection (ascending and descending stairs, running, walking, shuffle walking, sitting, standing, lying down -on right, left, back and stomach- , and miscellaneous movements).
- Achieved an accuracy of over 80% for respiratory condition classification (normal, hyperventillation, coughing, and others - laughing, talking, singing).
- Leveraged different regularisation techniques to improve validation and real-time performance.
- Supporting the background on IoT research, investigated the use of IoT devices in healthcare applications for the elderly.
Enhancing Sheep Detection Techniques with Deep Learning
- Collaborated in a team of three to enhance sheep counting accuracy using PyTorch.
- Identified challenges of existing literature in detecting occluded sheep and selected a more comprehensive dataset to improve performance in such scenarios.
- Implemented transfer learning techniques to adapt pre-trained models for sheep detection, such as SSD and Faster R-CNN.
- Utilized Nvidia GPU servers to accelerate model training.