I co-developed a testing framework enabling ROS-based systems to test any subset of their architecture, from individual nodes to complete systems.
This framework turned to be so effective that the company chose to open-source it. I’m currently working on the first public release.
Highlight: Speed Dependent Safety Zone Module @ Agrirobot
I implemented the prototype of a critical safety subsystem that dynamically calculates obstacle detection zones based on robot speed and size to ensure safe stopping distances.
Built the initial prototype from theoretical modeling through to a 2D dynamic simulator, accounting for kinematic effects. The end results is the calculation of detection areas that allow the robot to operate safely.
Took part in the core development as part of a 3-person team, building the first prototype for the autonomous tractor project. Implemented path-tracking, path-planning, and mission handling modules with basic hardware (depth camera, Arduino, GPS).
The internal demo was succesful and managed to convince upper management, who were skeptical that the project could be done in-house, to shift from subcontractors to in-house development, scaling the team from 3 to 15 developers.
The role involved extensive field testing to validate the system, which became the company’s main product with original implementations still in production use today.
Highlight: Dev-support engineer:
For about half a year, I joined the dev-support team troubleshoot the robots deployed on the field in USA. I worked entirely remotly and Oversaw operating issues, with a team of 3 handled with over 40 tickets/month.
Part of this role involved developing diagnostic tools for triaging operation faults, for it to be used both by technical and non-technical people. This tool ended being very crucial for everyday operation.
Main Project: Windowed Hierarchichal Cooperative A* Multi-Agent Planning.video, Code
Developed a multi-agent planner that uses a windowed hierarchical approach to solve cooperative pathfinding problems.
Uses a hierarchical approach to solve the problem, where a high-level planner assigns subgoals to agents and an agent-level planner solves the pathfinding problem.
To solve conflicts, the planner uses a windowed approach, where agents are allowed to replan their paths within a time frame if they collide with other agents.
Main Project: Epistemic Logic for Reinforcement Learning. Report
On the topic of RL: I’m currently studying how Reinforcement Learning works in Multi-Agent Systems, on a special course dedicated to the subject. Nothing too crazy yet, but I’m sure I’ll have something to share (hopefully).
Here are some blog posts related to the topics I’ve studied:
On the topic of Logic: I studied how epistemic logic could be used to help in the learning process of reinforcement learning agents. This was the final work of that class. The work is available here and was done together with Elle McFarlane a great AI engineer if you need one.