Connectionists: [Meetings] Call for Competition｜CIS-RAM 2026 Embodied AI Grand Challenge - World Model for Embodied Intelligence (EAGC –WM4EI 2026)

Wed Jun 3 11:18:49 EDT 2026

[cid:43a16852-6099-41a2-b6a7-409c99935ad4]

  1.  Introduction

EAGC 2026 is an international competition focused on evaluating the capabilities of world models for embodied AI. Unlike many traditional benchmarks, its core goal is to advance the evaluation of world models from "whether they can generate plausible futures" to "whether they can support reliable actions," assessing a model's ability to understand and act within continuous interaction and real-world tasks.

  1.  Core Philosophy and Objectives

According to the official information, EAGC 2026 aims to evaluate world model capabilities for embodied intelligence in standardized simulated environments. Its focus goes beyond perception-centric benchmarks, emphasizing executable environment understanding, closed-loop decision-making, exception recovery, and rapid task adaptation.
Key Question: Can the participating systems construct executable environment or task representations from continuous observations and then use those representations for planning, execution, replanning, and recovery?
Evaluation Focus: It is not only about the final task success rate. More importantly, it evaluates whether a system can construct and maintain a structured internal model, update it, expose it (through verifiable files), and use it when unexpected changes occur.

  1.  Challenge Tracks in Detail

The competition features two complementary tracks evaluating the core capabilities of world models in environment modeling and task transfer.
A）Track 1: Unseen Environment Exploration and Closed-Loop Task Execution (UEE-CLTE)
Objective: To evaluate a system's ability for autonomous exploration, world model construction, task planning, and exception recovery in partially unseen simulated indoor environments. The system must discover spatial topology, recognize key objects, model interactive states and affordances, generate task plans, and replan after encountering a controlled exception.
Key Rules: This is a 90-minute closed-loop challenge. Each official hidden episode is evaluated as a single continuous, closed-loop run. Unrestricted resets, repeated trial-and-error exploration, or manual intervention are not allowed. At most one controlled exception will be triggered per test, preserving at least one recoverable solution path.
B）Track 2: One-Shot Demonstration-Driven Rapid Adaptation (OSD-DRA)
Objective: To evaluate a system's ability to rapidly form a task model from a single standardized expert demonstration. The system must infer task variables, constraints, action preconditions, object-relation changes, and success criteria, then generalize to hidden, novel task instances under a limited attempt budget.
Key Rules: The system receives only one chance to observe the expert demonstration before attempting a new instance. A maximum of three official attempts are allowed per instance. The evaluation emphasizes whether the system extracts the key task structure rather than merely reproducing the observed trajectory.

  1.  Registration and Participation

According to the provided registration links: https://www.cis-ram.org//2026//EAGC_2026_Website_Upload.html
Registration Requirement: All participating teams must complete registration before submitting results. Teams may select Track 1, Track 2, or both tracks. Each team must designate one team leader as the primary contact.
Registration Links:
Track 1 Registration: https://www.CIS-RAM.org//2026//EAGC/WMC4E_UEE-CLTE/registration
Track 2 Registration: https://www.CIS-RAM.org//2026//EAGC/WMC4E_OSD-DRA/registration
Eligibility: The challenge is open to university teams, industry research teams, independent researchers, and student teams worldwide.
Important Note: Registration does not replace the final submission. Teams must still submit executable containers, model files, logs, and technical reports according to the official submission instructions.
Important Information and Resources
Baseline Code References:
Track 1 Baseline
https://github.com/google-research/pathdreamer
Track 2 Baseline
https://github.com/raktimgg/osvi-wm
Official Evaluation Rules:

  *   All official evaluations will be conducted within the official runtime stack.
  *   Internet access, online model APIs, access to hidden ground truth, manual remote control, and large-scale retraining are not allowed during the evaluation.
  *   Lightweight online adaptation within the official time and compute budget is allowed.

Important Dates:
May 8, 2026: Competition registration opens
May 19, 2026: Dataset and code released
May 30 – June 5, 2026: Results submission window open
June 5, 2026: Baseline paper (deadline or publication)
June 6 – June 9, 2026: Results testing window
June 15, 2026: Competition registration closes
June 16, 2026: Environment v1.0 finalized/frozen
June 17 – June 26, 2026: Qualification round submission
June 27 – July 3, 2026: Hidden validation testing
July 5, 2026: Finalist teams announced
July 8 – July 15, 2026: Final version submission (locked/frozen submission)
July 16 – July 24, 2026: Official final testing
July 25 – July 28, 2026: Log auditing, re-running, and reproducibility checks
July 29, 2026: Final rankings released to teams
Early August 2026: Results announced and on-site finalist presentations/defense

Awards
According to the official website, EAGC 2026 aims to recognize overall performance, track-specific excellence, reproducibility, and research contribution. Awards include:
Main Awards:
Grand Champion
Track 1 : First prize; Second prize; Third prize
Track 2 : First prize; Second prize; Third prize
Special Awards
•Best World Model Design Award
•Best Open-Source Contribution Award
•Best Technical Report Award
•Best Academic Innovation Award
7. Related Resources and Links
•Competition Official Website: https://www.cis-ram.org//2026//EAGC_2026_Website_Upload.html
•Official Email: CIS-RAM2026_EAGC at outlook.com
Track 1 Resource List
•ProcTHOR / ProcTHOR-10K: Primary public training environments and procedural generation reference benchmark.
•Habitat Challenge / ReplicaCAD: Reference benchmark for mobile manipulation and hidden-test protocols.
•ALFRED: Reference for language-instructed task templates and long-horizon task structures.
•BEHAVIOR-1K: Reference benchmark for challenging, long-horizon household interaction tasks.
•Recommended Combination: ProcTHOR + Habitat Challenge + ALFRED
Track 2 Resource List
•RLBench: Primary public training dataset and benchmark platform.
•MimicGen: Demonstration data augmentation and task variant generation tool.
•LIBERO: Benchmark for cross-object, cross-goal, and cross-layout transfer learning.
•ManiSkill2 & RoboCasa: Extended resources for object variation and potential future challenge versions.
•Recommended Combination: RLBench + MimicGen + LIBERO
8. Summary
Through these two tracks, the EAIGC26-WMC4EI challenge aims to evaluate whether world models can truly become the internal mechanism for an embodied agent to understand the world, rather than merely remaining at the surface level of "appearing to understand the world."
Therefore, this challenge seeks to advance the evaluation of world models from "whether they can generate plausible futures" to "whether they can generate reliable actions." By placing models within the context of continuous interaction, task feedback, and generalization testing, the challenge will provide the field of embodied AI with a more application-oriented benchmark for world models. This ensures that world models are no longer evaluated solely as representation or generation modules, but undergo systematic testing within the process of completing real-world tasks.

Competition organizers:
Dongyan Huang, Tsinghua Univ., CN
Yixing Gao Jilin University, CN
Xutao Sun Jilin University, CN
Hongxia Xie Jilin University, CN
Jifeng Hu Jilin University, CN
Yan Wu A*STAR, SG

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260603/cce811ba/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 200635 bytes
Desc: image.png
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260603/cce811ba/attachment.png>