Oceanus is a multi-agent RL environment where AI agents learn to recover ghost fishing nets from the ocean. 640,000 tonnes of ghost gear is lost annually — killing marine life for decades. This dashboard shows AI agents solving that problem in real-time.

There are 6 AI agents running simultaneously: 4 ASV drones physically cleaning nets on the ocean grid, and 2 policy agents negotiating a treaty to slow net spawning at the source. A chaos adversary injects storms, equipment failures, and policy crises to test the agents.

Each ASV card shows: battery level (depletes by 1 per move, red when <25%), last action (MOVE/CLEAN/BROADCAST/SCAN), reasoning text, current sector, position, and nets cleaned.

The Coordination Score bar rises when agents broadcast positions to each other (+15% per broadcast) and clean nets (+20% per clean). High coordination = agents working as a team.

The Policy Negotiation panel shows the treaty progress bar and timeline. When Port Authority proposes a treaty and Fleet Manager accepts, the spawn rate drops immediately.

The agents shown in DEMO mode were trained using GRPO (Group Relative Policy Optimization) — the same algorithm behind DeepSeek-R1. Model: Qwen2-0.5B-Instruct with 4-bit LoRA (2.1M trainable parameters), trained for 600 steps on Google Colab T4.

Before training: model outputs random text → reward −1.0 per step.
After training: model outputs {"intent": "clean"} when a net is visible → reward +8.0 per step.

Verifiable improvement: +9.0 reward points per step.

Need a fresh start? Reset the simulation below.

Enable real LLM agents powered by Llama3-8B via Groq. When active, each agent's decisions come from a real language model — you'll see thought bubbles showing the LLM's reasoning above each drone in real-time.

Get a free API key at console.groq.com (takes 30 seconds).

ENABLE LLM USE MOCK

Mode: Mock agents (heuristic)