A competitive evaluation platform for multi-agent reinforcement learning.

TL;DR

  • FastAPI-based evaluation server for multi-agent racing on the MetaDrive simulator
  • Students upload trained agents and compete against all opponents, with live leaderboard and match replays.
  • Built for UCLA CS260R (Reinforcement Learning) — Winter 2026 competition completed

Features

Automated Matching

Uploaded agent automatically races against all existing opponents. Results refresh on the leaderboard.

🏆

Live Leaderboard

Real-time rankings by win rate, ELO ratings, and route completion stats.

🎥

Match Replays

Bird's-eye view and 3D camera replay videos for every episode.

🚀

GPU-Accelerated Evaluation

Use GPU for faster model inference. Supports concurrent evaluation and multiple GPUs.

🔒

Token Authentication

Students use unique tokens to upload agent and review their own match history.

🛠

Admin Dashboard

Useful tools for user management, round-robin, and grade exports.

Race Tracks

4 custom-designed maps testing different driving skills

Circuit
Oval
Serpentine
Hairpin

Architecture

Step 1
Agent Upload
Step 2
Job Queue
Step 3
GPU Worker
Step 4
MetaDrive Sim
Step 5
Leaderboard

Competition Results

CS260R Reinforcement Learning — Winter 2026

Acknowledgments

Designed and developed by Matthew Leng and Haoyuan Cai, VAIL @ UCLA.

Built on the MetaDrive simulator.