Projects
Selected research and engineering projects.
Long-Horizon LEGO Stacking with ACT
Problem: Train a low-cost robot arm to stack multiple objects from demonstrations.
Approach: Collect teleoperation data using SO101, train ACT-style policies, and evaluate task success, smoothness, and recovery behavior.
Tools: LeRobot, PyTorch, OpenCV, RealSense, SO101
Status: In progress
PPO-LSTM for Partially Observable Box Pushing
Problem: Solve a POMDP where the agent has only local sensor information.
Approach: Use recurrent PPO with LSTM/GRU memory, reward shaping, and belief-state analysis.
Tools: PyTorch, Gymnasium, PPO, recurrent networks
Status: Active
Isaac Lab Bimanual Toy Arrangement
Problem: Arrange scattered toys into a target goal configuration using dexterous bimanual manipulation.
Approach: Build simulation scenes, define observations/actions/rewards, and compare imitation learning and RL-style policies.
Tools: Isaac Lab, PyTorch, robot simulation, dexterous hands
Status: Research planning
Real-Time Fraud Risk Decisioning Platform
Problem: Score financial transactions in real time and decide whether to approve, review, or block.
Approach: Combine supervised ML, graph features, streaming features, and bandit/RL-style decisioning.
Tools: Python, scikit-learn/PyTorch, FastAPI, databases, streaming design
Status: Planned