The same navigation stack, on legs
DeepRobotics' Lite3 quadruped explores a MuJoCo house and forest with the planner the Star Nav drone uses. EGO-Planner plans at body height; a tracker turns its setpoints into velocity commands for Lite3's own learned walking policy; a terrain filter lets the planner take ramps and low steps instead of treating them as walls.
How it works
The planning stack is the drone's, unchanged except for its parameters. Two pieces are new: a terrain filter between the lidar and the planner's map, and a tracker between the planner's setpoints and the walking policy.
Setpoint → joint torques
- Tracker. Velocity = setpoint velocity + 1.5 × position error, at most 0.8 m/s, rotated into the body frame. The robot walks in any direction, so it does not turn to face its path first. Turn rate = 1.5 × yaw error. It stands still within 3 cm and 3°.
- Walking policy. DeepRobotics' PPO policy from Lite3_rl_deploy, run unchanged. Its 45 inputs are body rates, gravity direction, the command, joint angles and velocities, and the last action. Its outputs are joint offsets from the standing pose.
- PD. The same gains as the upstream deployment, every physics step.
Lidar → walkable map
- The drone's map treats everything above the floor as an obstacle, ramps and steps included.
- The terrain filter keeps a 0.1 m grid of the lidar returns. Each cell records its lowest return, taken as the ground, and its highest return up to 0.55 m above that.
- Obstacle: more than 9 cm of height within the cell (walls, legs, boxes), or a ground 9 cm to 0.55 m above a neighbour's (box tops, platform edges).
- Free: ramps, low steps, and anything above the robot, so it walks under the table.
- Obstacle cells reach the planner as vertical columns, so it avoids them at any height.
What it can walk on
The policy is blind: it feels the ground through its legs but cannot see it. Measured in the house corridor, in both directions:
| Terrain | Result |
|---|---|
| 12° ramp | Climbs |
| 7 cm steps | Reliable, up and down |
| 8 cm steps | Sometimes stalls |
| 10 cm steps | Falls |
| Table top at 0.75 m | Walks under it |
| 15 cm box | Treated as an obstacle and walked around |
| Command | Speed reached (flat ground) |
|---|---|
| Forward 0.5 m/s | 0.49 m/s |
| Sideways 0.3 m/s | 0.24 m/s |
| Turn 0.5 rad/s | 0.45 rad/s |
Tested every run
lite3_smoke.test walks the full mission headless and fails on any
contact, fall or waypoint timeout.
| Run | Result |
|---|---|
| House, lidar (default) | 12/12 waypoints, 0 contacts, reached the platform (torso peak 0.54 m) |
Forest, lidar (world:=forest) | 12/12 waypoints, 0 contacts, all 107 nearby pillars in the map |
| Forest, depth camera | 12/12 waypoints, 0 contacts |
- The pose is the simulator's ground truth; there is no odometry or SLAM in the loop yet.
- The planner does not know the gait's limits: its 9 cm step threshold is above the 7 cm the policy climbs reliably.
- The terrain filter is for static worlds: cells are never cleared.
- EGO-Planner is a drone planner used here in one horizontal layer.
- No hardware branch yet; CI does not run these tests because the model is downloaded.
Quick start
Ubuntu 20.04 with ROS Noetic and the workspace built as for the drone (see the drone page). The Lite3 model and policy are downloaded at pinned commits with a checksum, not committed.
01Install the walking robot
./src/scripts/setup_mujoco_env.sh --lite3 # onnxruntime + model + policy
02Watch it walk
roslaunch star_nav_bringup exploration_lite3.launch use_rviz:=true # house, lidar
rostopic pub -1 /exploration/trigger geometry_msgs/PoseStamped "{}"
03Headless tests
rostest star_nav_bringup lite3_smoke.test # house, lidar
rostest star_nav_bringup lite3_smoke.test world:=forest # pillar forest
rostest star_nav_bringup lite3_smoke.test world:=forest sensor:=depth
Licensing and credits
The Lite3 model (deep_robotics_model) and walking policy (Lite3_rl_deploy) are DeepRobotics' work under BSD-3-Clause. They are downloaded at setup, not redistributed here. The planning stack is GPLv3 (EGO-Planner); the simulator, tracker and terrain filter are original MIT code. Details in docs/LICENSES.md.