Multi-Robot Navigation (db-CBS)

EAI’s multi-robot navigation component generates conflict-aware trajectories for ground robots in one scene and tracks them inside a single SimulatorSession. It uses db-CBS by default and lets you assign goals to multiple robots through either a visual panel or the command line.

The component does not require ROS or Nav2. Existing ROS/Nav2 workflows remain available as an independent navigation option.

Feature Overview

• Plan mutually conflict-free trajectories for multiple ground robots in one batch.

• Select robots and goals directly in the Isaac Sim viewport.

• Assign goals to part of the team from the command line, or cycle the full team through their initial positions.

• Monitor robot clearance during execution, hold the team when robots get too close, and replan from current positions.

• Start with a built-in three-robot Factory example containing Carter, Scout, and Go2.

The component currently requires num_envs=1. It excludes aerial types such as CF2X, Iris, and Pegasus by default and manages ground articulations only.

Runtime Flow

EAI selection + current robot poses + per-robot goals
        ↓
Occupancy map → obstacle boxes + planning frame
        ↓
db-CBS batch planning with inter-robot conflict constraints
        +
Ground robots without goals reserved as stationary obstacles
        ↓
dynoplan optimization → synchronized double-integrator trajectories
        ↓
Continuous clearance validation → per-frame velocity/position commands
        ↓
Team hold on close approach → background replan → resume execution

Initial planning and runtime replanning run on one background worker. While a plan is pending, compute_actions() returns hold commands so the native planner never blocks the Isaac Sim frame loop.

First Build

Prepare Isaac Lab’s env_isaaclab environment as described in the Installation Guide. The native db-CBS target also needs CMake, C++17, Boost, Eigen, FCL, yaml-cpp, and Crocoddyl. The build script resolves these libraries only from the active Conda environment, /opt/openrobots, /opt/ros/humble, and existing loader paths; it does not install system packages.

cd /path/to/eai-simulator
conda activate env_isaaclab
python -m pip install --no-deps crocoddyl==2.0.2
EAI_DBCBS_BUILD_JOBS=8 algorithm/multi_robot_navigation/build_native.sh

After a successful build, the script automatically prepares the motion primitives required by db-CBS. If they are not installed locally, it downloads them from the location provided by upstream db-CBS and verifies both file size and SHA-256. Existing files are verified and reused.

Interactive Navigation

Launch the three-robot Factory scene:

python -m algorithm.multi_robot_navigation.integration \
  --env dbcbs_slam_team --real-time

When neither --goal nor --exchange is supplied, the demo opens the EAI Multi-Robot Navigation panel:

1. Click a ground robot in the Isaac Sim viewport.

2. Click the floor or another pickable surface to assign a world-coordinate goal.

3. Repeat for every robot that should participate.

4. Click Start Navigation to submit all assigned goals as one batch.

A yellow ring marks the current selection. Per-robot colors show goals and remaining paths. Ground robots without goals stay stopped and reserve their current positions during planning.

Interactive mode needs a visible viewport and cannot be combined with --headless.

Command-Line Missions

Repeat --goal ROBOT:X,Y to assign only part of the team:

python -m algorithm.multi_robot_navigation.integration \
  --env dbcbs_slam_team \
  --goal carter_1:3.0,3.0 \
  --goal go2_1:-2.0,1.0 \
  --real-time

Cycle every ground robot to the next robot’s initial position:

python -m algorithm.multi_robot_navigation.integration \
  --env dbcbs_slam_team --exchange --real-time

Non-interactive missions can use --headless. --max-seconds controls the mission timeout and --hold-seconds controls how long the simulation remains open after completion; neither changes the native planner timeout.

Maps

from_session() selects a map under algorithm/multi_robot_navigation/maps/ from the environment JSON’s scene_key. The built-in scene keys are:

airs, desert, factory, garden, hospital, plane, warehouse

Pass a matching map explicitly when scene geometry changes or when using a custom scene:

python -m algorithm.multi_robot_navigation.integration \
  --env my_env \
  --map-yaml /absolute/path/to/map.yaml \
  --goal carter_1:2.0,-1.0 \
  --real-time

Maps use the common occupancy-grid YAML fields image, resolution, origin, occupied_thresh, free_thresh, and negate. The image and origin must match EAI world coordinates. The current db-CBS conversion requires a zero-yaw map origin.

Planning and Safety Semantics

Each robot has a planar bounding radius. Defaults are resolved by robot type and can be overridden by type or instance name through dbcbs_robot_radii. Planning also adds dbcbs_safety_margin:

effective radius = robot radius + safety margin
required pair separation = left effective radius + right effective radius

A replacement trajectory is installed only when every goal in the batch succeeds. The optimized result is also checked for pairwise clearance over linearly interpolated continuous segments. A result that violates effective radii is rejected instead of being sent to controllers.

During execution, when a mission robot approaches another ground robot, the whole team first receives hold commands. The component then replans the original mission goals from current poses on the background worker. A failed replan does not cancel the mission; goals are retained and retried after the configured interval.

For db-CBS, start_navigation() acknowledges planner submission rather than final planner success. Read state().planning and state().planning_error for the final outcome.

Integrating an Existing Session

from algorithm.multi_robot_navigation.eai_plugin import (
    EaiMultiRobotNavigationPlugin,
)
from simulator import SimulatorLaunchConfig, open_simulator_session

config = SimulatorLaunchConfig(
    env="dbcbs_slam_team",
    enable_ros_bridge_extension=False,
)

with open_simulator_session(config) as simulator:
    navigation = EaiMultiRobotNavigationPlugin.from_session(simulator)
    try:
        navigation.set_goal("carter_1", (3.0, -2.0))
        navigation.set_goal("go2_1", (-1.0, 4.0))
        navigation.start_navigation()

        while simulator.simulation_app.is_running():
            simulator.env.step(navigation.compute_actions())
            state = navigation.state()
            if state.planning_error:
                raise RuntimeError(state.planning_error)
    finally:
        navigation.close()

compute_actions() returns commands only for managed ground robots. A host that also controls aerial robots or other agents must merge commands from its other components before env.step(). Always call close() during host shutdown to cancel queued work and release the planner worker. A running native process remains bounded by dbcbs_planning_timeout.

Key Parameters

Parameter	Default	Purpose
planner_backend	"dbcbs"	Use the default db-CBS backend; "global" is retained only for the legacy compatibility path.
dbcbs_planning_timeout	60.0	Maximum seconds for one native plan or replan.
dbcbs_robot_radii	None	Override planar radii by robot type or instance name.
dbcbs_safety_margin	0.10	Safety margin added to each robot radius, in metres.
dbcbs_coarsen_factor	4	Occupancy-grid downsampling factor before obstacle-box conversion.
dbcbs_replan_clearance	0.25	Extra distance beyond hard clearance that triggers an early replan.
dbcbs_replan_retry_interval	0.50	Minimum retry delay after a failed replan, in seconds.
exclude_aerial	True	Exclude known aerial robot types from the managed navigation team.

Current Limitations

• The planning model is a planar double integrator with a fixed 0.1 s step. Height and full rigid-body dynamics are outside the db-CBS problem.

• The occupancy map is static. The component reacts to robot-to-robot proximity but does not update obstacle cells from sensors.

• A custom scene needs a map aligned with its world coordinates and geometry.

• Only one parallel environment is supported; num_envs > 1 is rejected.

Sources and Acknowledgments

This integration is based on IMRCLab/db-CBS (MIT).