Project Flock

Simulation Dashboard

Inspiration

Robot fleets are usually coordinated through direct communication or a central scheduler. We wanted to explore a different model: what if Cockroachdb itself could become the fleet's shared memory and coordination layer?

What it does

Project Flock simulates a disaster-response fleet of scouts, lifters, and medics that find victims, clear debris, and deliver aid. Robots cannot communicate directly. Instead, they coordinate through CockroachDB agentic memory, claiming tasks, sharing observations, handing off work, and recovering when robots or database nodes fail.

How we built it

We built the simulation with Python, FastAPI, WebSockets, Three.js, CockroachDB, and AWS Bedrock. CockroachDB stores four forms of fleet memory: working, episodic, provenance, and semantic memory. Serializable transactions and leases coordinate task ownership, changefeeds propagate operator interventions, and Claude with Titan embeddings handles planning and cross-mission learning.

Challenges we ran into

The hardest part was making coordination reliable without accidentally rebuilding a central scheduler. We had to handle concurrent task claims, duplicate observations, robot failures, database failures, and persistent learning while keeping every decision traceable.

Accomplishments that we're proud of

We built a decentralized fleet that can continue a rescue mission through a database-node failure, automatically recover abandoned work, learn reusable tactics between missions, and explain why robots made particular decisions. We also built both 2D and 3D mission visualizations and a test suite with 896 tests.

What we learned

We learned that Cockroachdb can be potentially used for more than storage. With transactions, leases, vector search, historical state, and changefeeds, it can become the coordination and memory substrate for autonomous agents used in robot fleets.

What's next for Project Flock

We want to test the architecture across larger fleets, more complex environments, and eventually real robotic systems to understand how persistent fleet intelligence could generalize beyond disaster response.