Documentation Index
Fetch the complete documentation index at: https://mintlify.com/jackvice/RoboTerrain/llms.txt
Use this file to discover all available pages before exploring further.
RoboTerrain integrates the ROS 2 Nav2 navigation stack to provide a classical planning baseline for comparison against reinforcement learning policies trained with Stable-Baselines3. Two fully-configured sensor variants are provided — one built around the Leo Rover’s 360° GPU LIDAR and another built around a depth camera point cloud — allowing head-to-head evaluation of Nav2 against RL agents across the same DUnE benchmark worlds. A third general-purpose parameter file is included for physical rover deployments with a pre-built map.
Nav2 is used as a comparison baseline in the DUnE benchmark. Running these configurations alongside trained RL agents lets you directly compare classical motion planning with learned navigation policies under identical terrain and obstacle conditions.
Configuration Files
Three Nav2 parameter files ship with RoboTerrain, each targeting a different sensor configuration or deployment context.
| File | Sensor | Use Case |
|---|
leo_nav2_lidar_params.yaml | 360° GPU LIDAR (/scan) | Simulation baseline — LIDAR-based obstacle avoidance |
leo_nav2_depth_params.yaml | Depth camera (/depth_camera/points) | Simulation baseline — depth camera obstacle avoidance |
nav2_params.yaml | LIDAR + MPPI controller | Physical rover with pre-built map and filtered odometry |
ros2_ws/src/roverrobotics_ros2/roverrobotics_gazebo/config/ # simulation configs
ros2_ws/src/roverrobotics_ros2/roverrobotics_driver/config/ # physical rover config
Key Nav2 Parameters
BT Navigator (bt_navigator)
The Behavior Tree navigator drives top-level mission execution. Both simulation configs use the odom frame as the global frame — no pre-built map or AMCL localization is required.
# leo_nav2_lidar_params.yaml
bt_navigator:
ros__parameters:
use_sim_time: true
global_frame: odom # No map/AMCL — odom is the reference frame
robot_base_frame: base_footprint
odom_topic: /odom_ground_truth # Ground-truth odometry from Gazebo
bt_loop_duration: 10 # ms per BT tick
default_server_timeout: 20
wait_for_service_timeout: 1000
default_nav_to_pose_bt_xml: /opt/ros/humble/share/nav2_bt_navigator/behavior_trees/navigate_to_pose_w_replanning_and_recovery.xml
nav2_params.yaml) switches to the map frame with filtered odometry:
# nav2_params.yaml
bt_navigator:
ros__parameters:
use_sim_time: False
global_frame: map
robot_base_frame: base_link
odom_topic: /odometry/filtered
Controller Server (controller_server)
The LIDAR config uses DWB (Dynamic Window Approach) with tuned velocity limits for the Leo Rover. The physical rover config replaces DWB with MPPI (Model Predictive Path Integral) for smoother trajectory tracking.
LIDAR Config — DWB Local Planner
# leo_nav2_lidar_params.yaml
controller_server:
ros__parameters:
use_sim_time: true
controller_frequency: 15.0
min_x_velocity_threshold: 0.001
min_y_velocity_threshold: 0.5
min_theta_velocity_threshold: 0.001
failure_tolerance: 0.3
FollowPath:
plugin: "dwb_core::DWBLocalPlanner"
min_vel_x: -0.7 # Allows reverse
max_vel_x: 0.7 # m/s forward limit
max_vel_theta: 1.5 # rad/s rotation limit
max_speed_xy: 0.7
acc_lim_x: 1.0
acc_lim_theta: 2.5
decel_lim_x: -1.0
decel_lim_theta: -2.5
vx_samples: 20
vtheta_samples: 20
sim_time: 1.7 # Trajectory rollout horizon (seconds)
linear_granularity: 0.05
angular_granularity: 0.025
xy_goal_tolerance: 0.25
critics: ["RotateToGoal", "Oscillation", "BaseObstacle", "GoalAlign",
"PathAlign", "PathDist", "GoalDist"]
BaseObstacle.scale: 0.15
PathAlign.scale: 32.0
GoalAlign.scale: 24.0
PathDist.scale: 32.0
GoalDist.scale: 24.0
RotateToGoal.scale: 32.0
RotateToGoal.slowing_factor: 5.0
Physical Rover Config — MPPI Controller
# nav2_params.yaml
controller_server:
ros__parameters:
controller_frequency: 30.0
FollowPath:
plugin: "nav2_mppi_controller::MPPIController"
time_steps: 56
model_dt: 0.05
batch_size: 2000
vx_max: 0.5
vx_min: -0.35
wz_max: 1.9
temperature: 0.3
gamma: 0.015
motion_model: "DiffDrive"
critics: ["ConstraintCritic", "ObstaclesCritic", "GoalCritic",
"GoalAngleCritic", "PathAlignCritic", "PathFollowCritic",
"PathAngleCritic", "PreferForwardCritic"]
ObstaclesCritic:
repulsion_weight: 1.5
critical_weight: 20.0
collision_cost: 10000.0
collision_margin_distance: 0.1
inflation_radius: 0.55
cost_scaling_factor: 10.0
Planner Server (planner_server)
# leo_nav2_lidar_params.yaml — SmacPlanner2D (MOORE motion model)
planner_server:
ros__parameters:
use_sim_time: true
expected_planner_frequency: 20.0
planner_plugins: ["GridBased"]
GridBased:
plugin: "nav2_smac_planner/SmacPlanner2D"
tolerance: 0.5
allow_unknown: true
max_iterations: 1000000
motion_model_for_search: "MOORE"
cost_travel_multiplier: 2.0
# leo_nav2_depth_params.yaml / nav2_params.yaml — NavfnPlanner (Dijkstra/A*)
planner_server:
ros__parameters:
planner_plugins: ["GridBased"]
GridBased:
plugin: "nav2_navfn_planner/NavfnPlanner"
tolerance: 0.5
use_astar: false
allow_unknown: true
Costmap Parameters
Both local and global costmaps use a rolling window (no static map). Key sizing and robot radius values:
# leo_nav2_lidar_params.yaml — local costmap
local_costmap:
local_costmap:
ros__parameters:
rolling_window: true
width: 6 # metres
height: 6 # metres
resolution: 0.05 # 5 cm per cell
robot_radius: 0.4 # Leo Rover approximate footprint radius (m)
plugins: ["obstacle_layer", "inflation_layer"]
inflation_layer:
cost_scaling_factor: 5.0
inflation_radius: 0.4 # metres around obstacles
# leo_nav2_lidar_params.yaml — global costmap
global_costmap:
global_costmap:
ros__parameters:
rolling_window: true
width: 20
height: 20
resolution: 0.05
robot_radius: 0.3
inflation_layer:
cost_scaling_factor: 3.0
inflation_radius: 0.5
# nav2_params.yaml
local_costmap:
local_costmap:
ros__parameters:
width: 8
height: 8
footprint: "[[0.27305, 0.1778],[0.27305, -0.1778],[-0.27305, -0.1778],[-0.27305, 0.1778]]"
footprint_padding: 0.01
plugins: ["voxel_layer", "inflation_layer"]
inflation_layer:
cost_scaling_factor: 3.0
inflation_radius: 0.35
Velocity Smoother
# leo_nav2_lidar_params.yaml
velocity_smoother:
ros__parameters:
smoothing_frequency: 20.0
feedback: "OPEN_LOOP"
max_velocity: [0.7, 0.0, 0.7] # [vx, vy, vtheta]
min_velocity: [-0.7, 0.0, -0.7]
max_accel: [1.0, 0.0, 2.5]
max_decel: [-1.0, 0.0, -2.5]
odom_topic: "odom_ground_truth"
LIDAR vs Depth Camera Config
The two simulation parameter files differ primarily in their observation source, planner algorithm, and costmap update rates. The table below summarises the key differences:
| Parameter | leo_nav2_lidar_params.yaml | leo_nav2_depth_params.yaml |
|---|
| Sensor topic | /scan (LaserScan) | /depth_camera/points (PointCloud2) |
| Sensor frame | lidar_link | camera_link |
| Obstacle max range | 8.0 m (costmap), 10.0 m (raytrace) | 5.0 m |
| Obstacle min range | 0.15 m | 0.2 m |
| Min/max obstacle height | Up to 15.0 m (2D scan, no z filter) | 0.1 m – 2.0 m (3D filter) |
| Controller frequency | 15.0 Hz | 20.0 Hz |
| Local costmap update | 5.0 Hz | 10.0 Hz |
| Planner plugin | SmacPlanner2D (MOORE grid) | NavfnPlanner (Dijkstra) |
| Expected planner frequency | 20.0 Hz | 5.0 Hz |
| Local inflation radius | 0.4 m | 0.55 m |
| Global inflation radius | 0.5 m | 0.55 m |
| Inflation cost scaling | Local 5.0 / Global 3.0 | 2.0 (both) |
| Odom topic | /odom_ground_truth | /odom |
| Recovery behaviors | spin, backup, drive_on_heading, wait | spin, backup, wait |
The depth camera config does not include a velocity_smoother or waypoint_follower node in its lifecycle manager. Ensure your launch file accounts for this if you are mixing components from the two configs.
BaseObstacle critic is weighted significantly higher (0.15 vs 0.02 in the depth config), making it more conservative around detected obstacles. The depth config compensates with a tighter inflation_radius and height-filtered point clouds to avoid penalising distant ground returns.
Using Nav2 as a Baseline
Launch the LIDAR-based Nav2 stack for comparison against an RL policy:
# LIDAR-based Nav2 navigation
ros2 launch roverrobotics_gazebo leo_nav2_lidar_launch.py
# Depth camera-based Nav2 navigation
ros2 launch roverrobotics_gazebo leo_nav2_depth.launch.py
ros2 topic pub /goal_pose geometry_msgs/PoseStamped \
"{header: {frame_id: 'odom'}, pose: {position: {x: 5.0, y: 3.0, z: 0.0}, \
orientation: {w: 1.0}}}" --once
odom frame selected as the fixed frame.
To record metrics for DUnE benchmark comparison, echo the /navigate_to_pose/_action/status topic and log success/failure alongside the RL episode logs:
ros2 topic echo /navigate_to_pose/_action/feedback \
| grep -E "distance_remaining|navigation_time"
Both Nav2 configs run with use_sim_time: true and pull ground-truth odometry from Gazebo (/odom_ground_truth). This ensures a fair comparison against RL agents that also receive ground-truth pose during training. For real-world evaluation, switch use_sim_time to false and point odom_topic at your filtered odometry source.
