STATE-NAV: Stability-Aware Traversability Estimation for Bipedal Navigation on Rough Terrain

Update News!

Now ROS2 version available! Still stabilizing though. Currently preparing Docker environment. Now both versions (ROS1/ROS2) are available for multiprocessing with shared memory.

Overview

STATE-NAV is the first learning-based traversability estimation and navigation framework for humanoids on diverse rough terrain. It learns a stability-aware velocity-based traversability representation of terrain by carefully selecting a self-supervised locomotion signal for bipedal locomotion and integrates this knowledge into a hierarchical planning framework for safe and efficient navigation.

Practical sub-module use cases

• Get TravRRT for Terrain Navigation: A computationally efficient RRT* for terrain navigation with traversability. Theoretically, RRT* assumes Lipschitz-continuous costs, which break on rough terrain where traversability changes abruptly, and forcing large rewiring radius. TravRRT* avoids this by biased sampling toward high-traversability regions—much like how humans just ignore infeasible paths (you don’t even think about walking through a bush, do you?). (Why not A*? yaw angle makes the search 3D! High computation. And also, heuristics for computing h(x).)

• Get Bipedal Traversability for Humanoids: An off-the-shelf bipedal traversability map—insert an elevation map, receive a traversability map.

Why STATE-NAV?

• Carefully selected self-supervising signal for bipedal locomotion: The first learning-based traversability estimation framework for bipedal locomotion. Uses Body-to-Stance-Foot Angle (BFSA), which has high correlation with bipedal fallover, as a traversability label for self-supervision. Unlike previous approaches that rely on traction or IMU signals (used for quadrupeds but not validated for bipedal instability), BFSA provides a more appropriate signal for bipedal locomotion.

• Velocity-based traversability: Robot-specific (environment-agnostic) cost formulation: Represents traversability not as a unitless cost or score, but as velocity with physical meaning. Defines the fastest velocity that a robot can traverse while maintaining instability (a robot-specific parameter) risk-sensitively within acceptable limits. This velocity representation replaces traditional path planning costs of $c_{ij} = \sum_{k \in M} (1 + w(1/t_k))^p \Delta l_k$ which depends on environment-specific hyperparameter $w$ that requires retuning for every environment.

• Safe learning-planning integration: Rather than end-to-end learning, leverages the learned model where it can do what models cannot and integrates the learned representation effectively within a hierarchical planning strategy—as a cost term in RRT* path planning and as a constraint in MPC local planning.

• Modular and extensible: Provides researchers with traversability maps and path plans for bipedal locomotion trained with Digit. Support for other robots will be added in future releases.

Key Features

• Traversability Estimation: TravFormer, a transformer-based neural network, predicts bipedal instability along with its uncertainty. Traversability is defined as a stability-aware command velocity: the fastest command that keeps predicted instability within a user-specified limit. The model supports risk-aware planning via Value at Risk (VaR) and operates directly from elevation maps.

• TravRRT Global Planning: A global planner that leverages the predicted stability-aware velocity to generate time-efficient and risk-sensitive paths. Consistently avoids unsafe terrain without manual weight tuning or environment-specific adjustments.

Citing

If you use STATE-NAV in your research, please cite:

@ARTICLE{11316382,
  author={Yoon, Ziwon and Zhu, Lawrence Y. and Lu, Jingxi and Gan, Lu and Zhao, Ye},
  journal={IEEE Robotics and Automation Letters}, 
  title={STATE-NAV: Stability-Aware Traversability Estimation for Bipedal Navigation on Rough Terrain}, 
  year={2025},
  volume={},
  number={},
  pages={1-8},
  keywords={Navigation;Planning;Robots;Estimation;Humanoid robots;Uncertainty;Stability criteria;Legged locomotion;Costs;Training;Humanoids;legged robots;traversability;navigation;planning;model predictive control;stability},
  doi={10.1109/LRA.2025.3648502}}

Quick instructions to run

ROS2 version will be supported soon.

Cloning Repositories

Clone the repo:

cd $(path_to_your_ros2_workspace)/src
git clone

Easy: Prebuilt Docker Image

The easiest way is downloading a prebuilt container from docker.io/twoleggedcoffeedrinker/state_nav:latest. If it is not available, you can choose to build the docker image from scratch as following:

If not: Manually building Docker container

Preliminaries

First, Install Docker: https://docs.docker.com/desktop/install/ubuntu/

Second, install nvidia container toolkit: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html

Building an Image

cd $(path_to_your_ros2_workspace)/src/state_nav/docker
docker build -f Dockerfile.x64 -t biped_nav_sim .
# You might need to change the image name in `runfrombuild.sh`

Opening a container

After building, open a container.

Caution* Make sure to change the line 3 in your runfrombuild.sh for path to your home directory and ROS2 workspace.

• When reopening container throws an error for XAUTH, delete your /tmp/.X11-unix and /tmp/.docker.xauth and then running ./runfrombuild would work. Running ./runfrombuild with sudo might potentially cause a problem for setting home directory.

sudo rm -rf /tmp/.X11-unix
sudo rm -rf /tmp/.docker.xauth

./runfrombuild.sh

Opening shells and running the Startup script

After opening an container, open an interactive shell in the container.

Run the following. You should be in the container's filesystem.

. $HOST_HOME_DIR/$(path_to_your_ros2_workspace)/src/state_nav/docker/startup.sh

This will build ROS2 package. Now you are good to go. Keep in mind that this build is only valid in this specific container. If you open a new container, you have to do the same thing again. Also, when you do ROS2 build in a docker container, it might conflict if you do ROS2 build outside of the container.

Running

Run the followings in separate shells.

(0-1) rosbag play $(path_to_your_catkin_workspace)/src/state_nav/ROS/1017_2.bag

# Plays back recorded sensor data (camera pointcloud, elevation map, path plan, etc.) from a rosbag file.
# It is recording of one of our outdoor experiments.
# Replace $(path_to_your_catkin_workspace) with your actual catkin workspace path.

(0-2) rosrun rviz rviz -d $(path_to_your_catkin_workspace)/src/state_nav/ROS/rviz_setting.rviz

# Launches RViz visualization tool with a pre-configured setup to visualize
# the traversability maps, planned paths, robot pose, and other ROS topics.
# Replace $(path_to_your_catkin_workspace) with your actual catkin workspace path.

(1) Modify executables/configs/planning_config.yaml as you want.

(2) ros2 run statenav_global traversability_estimation

# Launches the traversability estimation node that converts elevation map to traversability map.
# Computes stability-aware traversability maps using the TravFormer neural network.
# This node publishes costmaps that indicate safe navigation regions for the bipedal robot.

(3) ros2 run statenav_global global_planning

# Launches the global path planning node that uses TravRRT* to compute optimal paths
# based on the traversability maps. This node subscribes to costmaps and robot pose,
# then publishes planned paths for navigation.

(4) ros2 bag play $(path_to_your_ros2_workspace)/src/state_nav/ROS/1017_2.bag

# Plays back recorded sensor data (camera pointcloud, elevation map, path plan, etc.) from a rosbag file.
# It is recording of one of our outdoor experiments.
# Note: ROS1 bags need to be converted to ROS2 format first using ros1_bridge.
# Replace $(path_to_your_ros2_workspace) with your actual ROS2 workspace path.

(5) rviz2 -d $(path_to_your_ros2_workspace)/src/state_nav/ROS/rviz_setting.rviz

# Launches RViz2 visualization tool with a pre-configured setup to visualize
# the traversability maps, planned paths, robot pose, and other ROS2 topics.
# Note: The RViz config file may need to be updated for ROS2 compatibility.
# Replace $(path_to_your_ros2_workspace) with your actual ROS2 workspace path.

Acknowledgments

This project builds upon the following open-source works:

Docker and RViz Configuration

The Docker setup and RViz visualization configurations are adapted from elevation_mapping_cupy (MIT License).

RRT* Path Planning

The RRT* (Rapidly-exploring Random Tree Star) implementation is based on pypolo (MIT License), a Python library for Robotic Information Gathering.

We gratefully acknowledge the contributions of these projects to the robotics community.