International Conference on Metaverse Computing, Networking and Applications (MetaCom 2026)

Nov 6-10, 2026 · Xi'an, China.

http://www.metacomm.org/2026

MARC (Meta-Sejong AI Robotics Challenge) 2026 —
A Vision-Language-Action Grounding Student Challenge Workshop on Embodied Metaverse Robotics

https://marc-challenge.github.io/en/

This workshop evaluates embodied robotic autonomy in a metaverse digital-twin environment, where a robot takes natural-language commands and multi-view CCTV imagery as input and carries out Vision-Language-Action (VLA) grounding, mobile manipulation, and semantic destination routing. It is organized as a simulation-based student challenge; the paper presentations and live demonstrations by the finalist teams constitute the workshop's academic program.

Organization Chairs

• Program Chair, Jaeho Kim, Sejong University, Republic of Korea, kimjh@sejong.ac.kr
• Co-Chair, JaeSeung Song, Sejong University, Republic of Korea, jssong@sejong.ac.kr
• Co-Chair, Li Bai, Temple University, USA, li.bai@temple.edu
• Technical Chair, Taehyun Kim, Sejong University, Republic of Korea, thyun@sejong.ac.kr

Topics

MARC 2026 addresses embodied robots that understand natural-language commands and multi-view CCTV imagery to localize a target and then autonomously navigate and manipulate it, within a metaverse digital twin (a 3D model of the Sejong University campus). Participants implement an integrated chain of capabilities: natural-language command interpretation, visual grounding, mobile manipulation, and delivery of the target to its designated place.

Focus — Complementarity with the Main Conference

• Metaverse × Embodied AI: The workshop sits at the intersection of the conference's Computing and AI tracks, addressing embodied robotic autonomy that leverages a metaverse digital twin as a simulation training ground. Delivered as a hands-on student challenge, this robotics-in-metaverse focus complements the regular paper tracks.
• VLA Grounding: It goes beyond object detection to evaluate the integration of language, vision, and action—identifying the relevant target and grounding it into the appropriate action.

Timeliness

VLA and embodied AI have recently become central topics in the robotics and AI communities, and demand for reproducible, safe simulation-based evaluation environments is rising sharply. By adopting a standard ROS 2 interface on Isaac Sim 5.1.0 (LTS), this workshop secures sim-to-real transferability and offers direct relevance to the MetaCom community.

Competition Structure

The challenge runs in two stages. Stage 1 (VLA grounding) scores visual localization and reporting on six axes: camera selection, target-type and situation recognition, landmark identification, spatial-relation interpretation (relation words such as beside, on, around, and behind), anchor (relative) position, and target-coordinate attribution. Stage 2 (mobile manipulation) has the robot autonomously navigate to and secure the Stage 1 target, then transport it to the designated place (the item owner's location). Teams may enter Stage 1 alone or both stages together (Stage 1 + 2)—an option that lowers the entry barrier for teams for whom mobile manipulation is too demanding.

Technical Platform

The challenge runs on NVIDIA Isaac Sim 5.1.0 (LTS) with ROS 2 Humble on Ubuntu 22.04. Navigation and manipulation are commanded through standard ROS 2 messages. The final score is the sum of the Stage 1 (VLA grounding) and Stage 2 (securing the item and transporting it to the designated place) scores.

Illustrative Scenario

As a concrete instantiation of the capabilities above, we are preparing a campus-safety scenario. A campus control room observes multiple CCTV feeds to recognize situations—search-and-rescue, emergency/accident response, and lost-item retrieval—and issues a natural-language command; the robot interprets the situation and locates the target in the multi-view feeds (Stage 1), and—for a lost item—secures it and transports it to the designated place such as the item owner's location (Stage 2). Stage 1 targets are people in distress—scored by situation category (normal, accident, emergency, or abnormal behavior)—and lost items, disambiguated from visually similar distractors by spatial-relation words (beside, on, around, behind) and surrounding context rather than by appearance. People in distress are represented in simulation rather than by modeling real persons, which avoids privacy concerns.

Review Process

• Finalist selection: finalists are selected by combining the platform's automated quantitative evaluation (objective Stage 1/Stage 2 scores) with a review of the workshop papers describing each team's approach.
• Final ranking: the final ranking is determined by the finalist teams' on-site presentations and live demonstrations at MetaCom.

Resources Required

The workshop is a full-day session—live-demo rehearsal in the morning and finals in the afternoon—with approximately three to four on-site finalist teams, finalized later based on the submissions. The program comprises finalist paper-presentation and live-demonstration sessions and one invited keynote. Each finalist team must present on-site, in a 20-minute slot per team (a 15-minute talk plus 5-minute Q&A). Required facilities are a projector and screen for presentations and network/display for live demonstrations. Sejong University provides the simulation platform and evaluation servers, while teams develop in their own environments.

Important dates

Milestone	Date (tentative)	Note
Platform release · Registration	2026-07-06 – 07-31	Registration opens together with platform release
Result submission	2026-08-31	Code and paper submission
Finalist announcement	around 2026-09-18	Reflects quantitative + paper review
Finals (IEEE MetaCom 2026)	2026-11-09, Xi'an, China	On-site presentation + live demonstration

Previous Editions

MARC is a competition hosted by Sejong University; its first edition was held in 2025. The inaugural MARC 2025 ran a campus waste-sorting (recycling) mission based on object detection; eight teams applied, and two teams ultimately earned the honor of winning the first edition.

MARC 2026 builds on this as the second edition, extending the task to VLA grounding, and is being organized with the goal of establishing MARC as an annually held event.

Official website: https://marc-challenge.github.io/en/