MARC 2026 Developer Guide

Welcome to the MARC (MetaSejong AI Robot Challenge) 2026 developer guide.

MARC 2026 at a glance

MARC 2026 is a robot-simulation competition in which you build a single AI agent that understands a person’s spoken request, locates the target in campus CCTV footage, and — when asked — drives a robot to retrieve it. Everything runs in a digital twin of the Sejong University campus on Isaac Sim 5.1.0, and the mission centers on VLA (Vision-Language-Action) grounding together with a SAR (Search-and-Rescue) perception task. The contest has two stages: Stage 1 — find and report the location (coordinate submission; required of every team) and Stage 2 — retrieve the object with the robot (navigation + manipulation; optional). For the full challenge story, rules, schedule, and organizers, see the competition homepage.

What this guide covers

This guide is your hands-on reference for development and submission. It walks you through the whole journey in order — from a clean machine with nothing installed, to building and running your own agent, to finally submitting it to the competition. Here is how each page helps you along the way:

  • Getting Started — the first page to read. It checks the hardware and tools you need (Ubuntu, an NVIDIA GPU, Docker, ROS 2, and so on), then takes the shortest path to downloading the starter kit, bringing the platform up, and watching the demo agent move — in about 30 minutes. Start here if you just want to see something running first.

  • Technical Guide — once the demo runs, this is where you add your own “intelligence” on top of it. It explains the participant SDK (marc_sdk) you use to talk to the platform, the structure of the baseline agent you build from, and the manipulation learning kit for practising how the robot arm picks objects up. It points out exactly where to swap in your own implementation.

  • API Reference — the reference for the contract by which the robot and sensors actually communicate: the commands that drive the robot, the sensor inputs (CCTV, cameras, lidar), and the coordinate frames, message formats, and units — all laid out in tables. Open this when you need to know “what value do I send, where, and in what format?”

  • Submission Guide — how to score your agent on your own machine to see how well it does, and how to prepare your submission in the right format and follow the submission procedure. It also covers how to read your score and per-round details after a run, and roughly how scoring works (the methodology).

  • FAQ — the page to check first when you get stuck. It gathers fixes for the problems teams hit most often (troubleshooting), along with the notices you must read before you take part.

For the competition rules, schedule, evaluation policy, and organizers, see the competition homepage.

Who this is for

Developers on participating teams. Working knowledge of Python, ROS 2, and Docker is assumed. Robotics/ML background helps but the baseline agent gives you a running start.

Important notices (read first)

Two required notices

  1. The evaluation runtime has no internet access. You may use the internet at build time (to bake weights and dependencies into your image), but at runtime external network access, public APIs, and downloads are prohibited.

  2. The competition background location may change. The only publicly distributed background USD is the practice scene (chungmu); the actual competition may use a different background.

See FAQ → Notices for the full text.

Contact

For questions about the challenge, the platform, or this guide, email marc2026@iotcoss.ac.kr.