LHNav: Long-Horizon Humanoid Navigation via Multirate Coordination and Physics-Grounded Action Data

Xueyang Wang, Zehong Shen, Dingkun Wei, Huaijin Pi, Changpeng Wang, Xing Zhu, Hujun Bao, Yujun Shen, Xiaowei Zhou

November 2026

Real Unitree G1 on the three-destination reading-notes task, with the planning agent state overlaid.

TL;DR A three-layer asynchronous system for multi-destination humanoid navigation from language: an LLM planner, a VLA that predicts whole-body motion tokens, and a frozen SONIC controller. It reaches 75.6% success on 45 long-horizon simulated tasks, and a Unitree G1 completes a three-destination household task in 4 of 5 trials.

What they do

A humanoid gets one long language instruction that implies several destinations in order (bring the notes to the bed, then leave them on the desk) and has to walk to each of them through a home. LHNav splits the job into three layers running at different rates, and builds a simulation data engine that produces physically executed walking demonstrations to train the middle layer. The navigation policy is trained only on single-goal demonstrations.

How it works

Brain. A planning agent (GPT-6 or Qwen3.8-27B) keeps an explicit plan of completed, active and pending goals, chosen from a catalog of destination images. A separate Qwen VLM checks arrival at up to 1 Hz; the next goal is dispatched only after verified arrival.
Cerebellum. A VLA fine-tuned from GR00T N1.7 takes the instruction, target image, current RGB and proprioception and jointly predicts 40-step chunks of SONIC motion tokens and auxiliary waypoints at about 8 Hz. A signed-distance penalty pushes the waypoints into traversable space.
Brainstem. The frozen SONIC decoder turns motion tokens into joint targets at 50 Hz.
Data engine. A* routes in 100 InteriorGS Gaussian-splat scenes with movable mesh obstacles, ARDY whole-body reference motions along them, and P-SONIC (a SONIC encoder fine-tuned with privileged root-position and velocity inputs) tracking them in MuJoCo. This yields 185,401 demonstrations with rendered first-person views.

Results

Single-goal, 522 episodes in 20 held-out scenes: 52.3% success within 0.5 m and 0 falls, against 36.0% and 7 falls for OmniNav and 18.6% and 91 falls for DualVLN.
Long-horizon, 45 simulated tasks with at least three ordered goals: 75.6% task success; 0% without task management and 55.6% without arrival verification.
P-SONIC cuts mean endpoint error on 10 m routes from 48.65 cm to 1.76 cm.
Real Unitree G1: 5/5 on each of five household destinations, 14/15, 14/15 and 15/15 on three targets across reconfigured layouts, and 4/5 on the three-destination reading-notes task.

Limitations

The paper states no limitations. From its setup: destinations come from a given catalog of reference images, real-robot tests follow scene-specific training, the long-horizon simulation tasks use scenes from the training split, and the real long-horizon test is one instruction with five trials. Navigation followed by manipulation is shown only as a qualitative toy-dog pushing example.

Videos

Real Unitree G1 walking to a household destination.

Long-horizon navigation in simulation, with the planning agent state overlaid.

Figures

From Figure 3: The asynchronous three-layer navigation system. The Brain uses task management to maintain an explicit... — From Figure 3: The asynchronous three-layer navigation system. The Brain uses task management to maintain an explicit plan and revise pending goals, with goal-completion feedback from an arrival-verification VLM at up to 1 Hz. The Cerebellum takes an instruction, a target image, current RGB and proprioception and jointly predicts motion-token chunks and auxiliary waypoints at approximately 8 Hz. The Brainstem’s frozen SONIC controller turns the latest chunk into joint targets at 50 Hz.

From Figure 4: Example of navigation data generation for “Go to the refrigerator”: scene annotation, route constructi... — From Figure 4: Example of navigation data generation for “Go to the refrigerator”: scene annotation, route construction, motion generation and the collected data.

From Figure 5: Reference tracking of SONIC and P-SONIC against the ARDY reference on the route in Figure 4. Motion ru... — From Figure 5: Reference tracking of SONIC and P-SONIC against the ARDY reference on the route in Figure 4. Motion runs from upper right to lower left; insets enlarge the endpoints.

From Figure 10: Real-robot toy-dog navigation across three reconfigured layouts. Each row shows the robot approaching... — From Figure 10: Real-robot toy-dog navigation across three reconfigured layouts. Each row shows the robot approaching the toy dog in one layout.

Abstract

We study humanoid navigation across multiple destinations under long-horizon language instructions. This task is challenging because robots must maintain progress across successive goals while continuously producing natural, stable whole-body motion through confined spaces. However, task decomposition in existing methods does not by itself ensure progress throughout execution, while waypoint-based navigation does not guarantee natural, route-faithful whole-body motion. Our key idea is to organize navigation into three complementary roles operating at different time scales: a Brain for task reasoning, a Cerebellum for local motion generation, and a Brainstem for whole-body execution. We realize this idea in LHNav, an asynchronous system that uses progress and arrival feedback to coordinate navigation goals, VLA motion-token prediction, and whole-body control. To train the Cerebellum for natural, route-faithful motion, we develop a physics-grounded data engine that improves physical tracking of motion references along collision-aware routes and pairs the resulting demonstrations with rendered visual observations. Together, these designs allow task planning to proceed alongside continuous movement and extend single-goal navigation to successive destinations with natural, route-faithful whole-body motion. In simulation, our navigation VLA outperforms OmniNav and DualVLN in single-goal success and motion smoothness, with fewer collisions and falls. Using a VLA trained only on single-goal demonstrations, LHNav achieves a 75.56% success rate on long-horizon navigation tasks. Deployment on the Unitree G1 further demonstrates single-goal navigation and zero-shot long-horizon task execution in real-world environments, including layouts with repositioned objects. Project website: https://robbyant-research.github.io/LHNav/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong