Skip to content

Challenge Participation

Official HA-VLN repository: https://github.com/JostarXiong/HA-VLN

This guide provides a quick overview of how to participate in the RoboWorld 2026 Track 2 (HA-VLN) Challenge. For full details, see the Challenge Overview and Submission Format.

Quick Start Checklist

1. Environment Setup

  • [ ] Pull pre-built Docker image (ghcr.io/jostarxiong/havln-challenge-2026@sha256:78a62cd176d2fd7d0e2825f4cb5be2488ebc5f1a354649b7b4f536a98f1054f4) or set up Native Conda (Py3.8 / CUDA 11.8)
  • [ ] Run verification tests (python scripts/demo.py --scan 1LXtFkjw3qL --headless)

2. Data Preparation

  • [ ] Download HA-R2R dataset & HAPS 2.0 human motion models via Hugging Face (hf download fly1113/HA-VLN)
  • [ ] Download Matterport3D scenes (download_mp.py + unzip into Data/scene_datasets)
  • [ ] Verify dataset layout

3. Agent Integration & Action Recording

  • [ ] Configure your agent for HA-VLN environment (HAVLNCE_task.yaml)
  • [ ] Record discrete string actions during inference
  • [ ] Test on val_seen (778 episodes) and val_unseen (1,839 episodes)

4. Submission

  • [ ] Export val_seen.json and val_unseen.json with format_version: 1
  • [ ] Create submission.zip containing both JSON files
  • [ ] Validate locally via Docker container: docker run --rm -v "$(pwd):/workspace" "$IMAGE" havln-validate /workspace/submission.zip
  • [ ] Submit to CodaBench

Action Vocabulary

The simulator environment evaluates six discrete string actions:

Action Description
"STOP" End episode (must be the final action)
"MOVE_FORWARD" Advance agent by 0.25m
"TURN_LEFT" Rotate heading left by 15°
"TURN_RIGHT" Rotate heading right by 15°
"LOOK_UP" Pitch sensor upward by 30°
"LOOK_DOWN" Pitch sensor downward by 30°

Recording Actions during Inference

# During agent inference, record actions as strings from the vocabulary:
action_names = tuple(config.TASK_CONFIG.TASK.POSSIBLE_ACTIONS)
action_traces = {str(ep.episode_id): [] for ep in current_episodes}

# In your step loop:
for episode, action_id in zip(current_episodes, actions):
    action_traces[str(episode.episode_id)].append(action_names[action_id])

Export each split to JSON (val_seen.json, val_unseen.json):

with open(f"{split}.json", "w", encoding="utf-8") as f:
    json.dump({
        "format_version": 1,
        "split": split,
        "episodes": [
            {"episode_id": ep_id, "actions": action_traces[ep_id]}
            for ep_id in sorted(action_traces)
        ]
    }, f, indent=2)

Package the submission:

zip -j submission.zip val_seen.json val_unseen.json

# Validate via Docker container (or 'havln-validate submission.zip' inside container):
IMAGE=ghcr.io/jostarxiong/havln-challenge-2026@sha256:78a62cd176d2fd7d0e2825f4cb5be2488ebc5f1a354649b7b4f536a98f1054f4
docker run --rm -v "$(pwd):/workspace" "$IMAGE" havln-validate /workspace/submission.zip