HumanoidVLN: Advancing humanoid robot navigation with physics-grounded AI
New simulator HumanoidVLN tests AI navigation on real-world humanoid robot bodies like Unitree G1 and H1.
HumanoidVLN is a physics-grounded simulator and benchmark for vision-language navigation across diverse humanoid embodiments, built on NVIDIA Isaac Sim. It supports four robot configurations—Unitree G1, Unitree H1, Internal-A, and Internal-B—with 10–12 lower-body degrees of freedom and heights from 1.17m to 1.80m, using a hierarchical control stack that pairs a reinforcement learning locomotion policy with interchangeable PD or MPC path trackers. The benchmark includes 933 collision-aware reference episodes, each with one fine-grained instruction and three coarse-grained stylistic variants. Across four models and four embodiments, JanusVLN achieves the highest mean success rate of 43.55% and an nDTW of 48.38. In a 20-episode sim-to-real pilot with DualVLN and the Unitree G1, navigation errors correlate strongly with real-world results (r=0.935), with a mean absolute difference of 0.68m and a mean trajectory similarity of 0.782 (±0.188) nDTW. Code, benchmark, and data are set to be released upon acceptance.
- HumanoidVLN uses NVIDIA Isaac Sim to simulate navigation for humanoid robots like Unitree G1 and H1, addressing bipedal physics and camera motion distortions.
- Benchmark includes 933 collision-aware episodes with stylistic variants; JanusVLN leads with 43.55% success rate and 48.38 nDTW across four robot models.
- Sim-to-real pilot shows high correlation (r=0.935) with 0.68m mean error and 0.782 trajectory similarity, validating simulator realism.
Why It Matters
HumanoidVLN bridges the gap between simulation and real-world humanoid robotics, enabling reliable testing of AI navigation models before deployment.