The RoboWorld Challenge 2026

🌍 Challenge Overview

This track evaluates an embodied agent's ability to perform socially compliant instruction following in photorealistic 3D environments populated with dynamic human agents, without relying on static-world assumptions:

• Human-Referenced Grounding: Interpret fine-grained instructions that reference both scene landmarks and ongoing human activities.
• Social Norm Compliance: Respect personal space, proactively avoid human collisions, and navigate safely around moving bystanders.
• Realistic Benchmarking: Navigate indoor and outdoor areas with multi-human interactions in the HA-R2R and HAPS 2.0 datasets.

We welcome navigation policies, waypoint and topological planners, world models, and VLA approaches that advance safe, human-centered navigation in shared spaces.

🎯 Task Definition

See the participant README for the action-sequence format and optional local replay instructions.

📅 Competition Details

• Event: RoboWorld Challenge 2026, Track 2.
• Joint workshop: RoboPAD Workshop at NeurIPS 2026.
• Registration: register through the Google Form (registration opens Oct 08, 2026) for leaderboard, certificate, and award eligibility.
• Submission platform: CodaBench — HA-VLN.
• Submission limits: five per day and 100 per phase; the best score is retained.

🏆 Awards & Recognition

Final rankings and awards are determined entirely by Phase 2 test evaluation.

🗂️ Phases

Both phases use the same six-action JSON contract and Score. Results are not combined.

🗓️ Timeline

Follow the competition page for any schedule updates.

📊 Dataset

The task combines HA-R2R route instructions with HAPS 2.0 human activity across indoor and outdoor areas of Matterport3D scenes. Matterport3D scenes require separate access under their own license.

HA-R2R contains 16,844 instructions across 90 scenes:

HAPS 2.0 provides 910 human models and 486 motion sequences of 120 frames each. Phase 1 submissions must cover all 778 val_seen and 1,839 val_unseen episodes; Phase 2 provides test inputs for final evaluation while reference trajectories and collision annotations remain withheld. Use the official episode manifests for identifiers; data preparation and simulator details are in the participant README.

🎬 Dataset Examples

These released HA-VLN 2.0 clips illustrate dynamic human activities and multi-human interactions in photorealistic scenes.

Released HA-VLN 2.0 media from the project page and code repository. Explore more scenes and activities on the HA-VLN 2.0 demo page.

🎬 Navigation Visualization

Two released HA-VLN 2.0 examples show how human activity becomes part of a navigation instruction.

Videos and instructions from the HA-VLN 2.0 repository.

🧠 CMA Baseline

We provide HA-VLN-CMA, a cross-modal attention navigation model from the released HA-VLN 2.0 project, as a reference baseline. Participants may use any method that exports legal action sequences. The original code, paper, and participant README provide implementation details and validation results.

📏 Evaluation

Higher full-precision Score ranks first. Ties use higher SR, lower NE, lower CR, lower TCR, then earlier submission time. See the participant README for metric definitions.

🔗 Contact and Resources

For technical questions, open an issue in the challenge repository. For competition inquiries, email roboworld2026@gmail.com.

For the complete data → CMA → action ZIP → submission workflow, follow the participant Quick Start.

❓ Frequently Asked Questions

1. Must I use CMA?
No. Any method may participate if it exports valid official action sequences.

2. What do I upload?
Only a ZIP containing the required JSON predictions. Awarded teams may later be asked for method, code, or model details to verify their results and contribute to a technical report.

3. Are look actions valid?
Yes. LOOK_UP and LOOK_DOWN are official actions and count toward the action budget.

4. Where are the datasets?
See the participant README. Request Matterport3D access through its official site.

5. Which replay determines my score?
Only the CodaBench replay of the submitted ZIP determines the official score. Local replay is for development.

6. What if my submission is rejected?
Use the format checker in the CodaBench starting kit and consult the Submission & Evaluation page for validation and scoring-log explanations.

📄 License and Terms

HA-VLN 2.0, HA-R2R, HAPS2.0, source media, pretrained weights, and Matterport3D retain their upstream licenses. Participation is governed by the Terms displayed on the official CodaBench competition.

📚 Citation

@inproceedings{dong2026havln,
  author    = {Dong, Yifei and Wu, Fengyi and He, Qi and Kong, Lingdong and Li, Heng and Li, Minghan and Cheng, Zebang and Zhou, Yuxuan and Sun, Jingdong and Dai, Qi and Hauptmann, Alexander G. and Cheng, Zhi-Qi},
  title     = {{HA-VLN 2.0: An Open Benchmark and Leaderboard for Human-Aware Navigation in Discrete and Continuous Environments with Dynamic Multi-Human Interactions}},
  booktitle = {2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
}

@misc{roboworld2026track2,
  title={Track 2 | HA-VLN: Human-Aware Vision-and-Language Navigation},
  author={RoboWorld Challenge 2026 Organizers},
  year={2026},
  howpublished={https://roboworld2026.github.io/track2}
}