ML Robotics for Elderly Assistance

A machine learning robotics project co-developed with two Harvard Summer School peers, targeting elderly assistance with daily household tasks. Combined computer vision for object recognition with a task planning and manipulation pipeline. Written up and submitted as a collaborative research project.

Build a practical robotic assistance system around the perception and manipulation problems involved in household chores. Tasks that are obvious to humans but poorly specified for robots.

• Three-person team across different backgrounds, which required clear task decomposition and parallel workstreams
• Limited physical hardware access. System architecture had to be testable primarily in simulation.

• Vision first: reliable object recognition was the prerequisite for any manipulation pipeline, so perception was developed and validated independently before anything else
• Modular architecture separating perception, planning, and actuation. Each layer testable without the full stack running.

1. 01

   Object detection and classification using trained CNN models

2. 02

   Task planning layer mapping recognized objects and scene state to chore sequences

3. 03

   Manipulation pipeline for common household object interactions

4. 04

   Integration and testing in simulated environment

• Real-world robustness in perception is much harder than benchmark accuracy suggests. Lighting variation alone breaks classifiers that looked solid on test data.