HumanoidVLN Adds Physics to Robot Navigation Tests

Single-source brief: An arXiv paper describes a simulator benchmark, not a deployed robot product.

What Changed

The HumanoidVLN team introduced a physics-based test system for humanoid navigation.

The system targets vision-language navigation, or following spoken or written route instructions.

Existing tests often miss problems unique to walking robots, the authors write.

Biped motion adds physical limits. Robot bodies also differ across platforms.

Camera views can shift as a robot walks. That can distort what navigation software sees.

HumanoidVLN runs on NVIDIA Isaac Sim. It supports different humanoid setups through a shared control stack.

The paper demonstrates four robots: Unitree G1, Unitree H1, Internal-A, and Internal-B.

Those robots have 10 to 12 lower-body degrees of freedom. Their heights range from 1.17 to 1.80 meters.

Benchmark Results

The authors report 933 collision-aware navigation episodes. Each has one detailed instruction and three broader writing styles.

They tested four navigation models across four robot bodies.

JanusVLN had the best reported average result. It reached 43.55% success and 48.38 nDTW.

That means most benchmark runs still did not count as successful.

Deployment Reality

This is a research simulator and benchmark. It is not evidence of broad real-world deployment.

The team also ran a 20-episode sim-to-real pilot. It used DualVLN on a Unitree G1.

The reported navigation errors showed a strong correlation between simulation and real runs.

Important unknowns remain. No independent confirmation was supplied.

The paper says code, data, and the benchmark will be released upon acceptance.