ON OmniNavBench
Beyond Isolation

OmniNavBench

A unified benchmark for general-purpose embodied navigation. We evaluate cross-skill coordination across VLN, PointNav, ObjectNav, HumanFollow, SocialNav, and EQA, together with cross-embodiment generalization on wheeled, quadrupedal, and humanoid robots.

GitHub Repo Hugging Face
Abstract

A Unified Platform for General-Purpose Navigation

Compositional Complexity

Composite instructions interleave sub-tasks from PointNav, VLN, ObjectNav, HumanFollow, SocialNav, and EQA, compelling agents to seamlessly transition between exploration, interaction, and social compliance within a single unified episode.

Morphological Universality

A high-fidelity Isaac Sim platform evaluates wheeled (NVIDIA Carter), quadrupedal (Unitree Aliengo), and humanoid (Unitree H1) robots under a modular sensor interface with continuous physical dynamics.

Human Expert Demonstrations

Reference trajectories are collected via human teleoperation rather than shortest-path planners, preserving exploratory glances, hesitations, and anticipatory avoidance that characterize natural human navigation.

Submission Protocol

Offline Result Submission

Accepted Input

Upload a single .zip archive. The archive may use any internal directory layout, but the scoring inputs are per-episode JSON files such as final_episode_*.json.

Reported Metrics

Leaderboard columns include navigation and interaction metrics — SR, CSR, SGC, OSR, SPL, HFS, HFR, SII, EQA — recomputed offline against the private ground truth.

Publishing Workflow

Uploads are queued for asynchronous evaluation. Once a submission succeeds, you can publish it from My Submissions to make it appear on the public leaderboard.

Results

Leaderboard

Pick a robot category and click any metric header to sort.

# Method User Affiliation Variant {{ m }}{{ lbSort===m ? '▼' : '' }}
No published entries yet. Sign in to submit the first result archive.
Tasks & Metrics

What the leaderboard measures

SRSuccess Rate — fraction of fully successful episodes
CSRCompletion Success Rate — ratio of strictly completed sub-instructions
SGCSub-Goal Completion — continuous progress across sub-tasks
OSROracle Success Rate — success if the goal was ever reached
SPLSuccess weighted by Path Length — efficiency-weighted success
HFS / HFRHuman-Follow final success and following-phase coverage
SII ↓Social Intrusion Index (Hall's personal distance, 1.2 m) — lower is better
Account Required

Submit and manage results

Create an account to upload episode result archives, monitor evaluation jobs, and publish successful submissions to the public leaderboard.

{{ authMode === 'login' ? 'Sign in' : 'Create account' }}

{{ authError }}
Submit

Upload a results archive

Submissions must be .zip archives containing per-episode JSON files. Precomputed summary metrics are ignored — all leaderboard metrics are recomputed server-side from raw trajectories against the private ground truth.

Submit Results

{{ submitError }}

Click or drop a result .zip here

{{ submitForm.file.name }}
The archive should contain one episode JSON per test case.
Submission History

My Submissions

Successful submissions can be published to the public leaderboard.

{{ historyError }}
No submissions yet.

{{ s.method_name }} {{ statusLabel(s.status) }} Published

{{ s.filename }} · {{ new Date(s.created_at).toLocaleString() }}
{{ v.toFixed(4) }} {{ k }}
{{ s.error_message }}