MobilityGen Synthetic Data Generation

Purpose

Run MobilityGen two-phase SDG: record robot trajectories headlessly, then replay and render RGB/depth/segmentation/normal/pose outputs.

Prerequisites

• Built Isaac Sim ($ISAAC_SIM_DIR or _build/linux-x86_64/release).
• NVIDIA GPU with a current driver (nvidia-smi).
• Shell env contract from isaac-sim-orchestrator: $ISAAC_SIM_DIR, $ISAAC_LAB_DIR, $WORKSPACE_DIR.

Limitations

• Targets Isaac Sim 6 / Kit 110 unless a section states otherwise.
• Does not replace official NVIDIA documentation for unsupported edge cases.

Troubleshooting

Two-phase pipeline: record trajectories (physics, no rendering) → replay & render (sensors added).

Available Scripts

Running scripts

From agent runtimes that expose skill execution helpers, invoke helpers with run_script():

run_script("scripts/holonomic_robot_subclass.py", args=["--help"])

From a built Isaac Sim tree, run the same file with ./python.sh (Linux) or python.bat (Windows) from _build/*/release, or execute shell helpers directly when they do not require the simulator.

Read These Skills First

• navigation-primitives — OccupancyMap, A* planner, robot footprints (Spot Z=0.69), differential/holonomic kinematics, look-at chase cameras, shared gotchas. MobilityGen consumes this substrate; this skill assumes you know it.
• occupancy-map — produces the map.yaml consumed by OccupancyMap.from_ros_yaml
• data-collection-sim — sibling SDG path for static scenes with randomized object/camera poses (no robot trajectory)

When To Use This Skill (vs siblings)

Related Skills

• navigation-primitives — shared navigation substrate (read first)
• data-collection-sim — static-scene SDG sibling
• isaac-sim-sensor — sensor primitives (camera, LiDAR, IMU, contact)
• isaac-sim-robot-navigation — runtime navigation sibling
• isaac-sim-headless-deployment — --no-window headless launch and SimulationApp batch pattern

Environment

• Isaac Sim source tree: $ISAAC_SIM_DIR/source/ for source builds. The variable $ISAAC_SIM_SRC is a convenience alias for that path; declare it once at the top of your launcher (e.g. ISAAC_SIM_SRC="$ISAAC_SIM_DIR/source").
• Python launcher: $ISAAC_SIM_DIR/python.sh
• Replay script: $ISAAC_SIM_SRC/standalone_examples/replicator/mobility_gen/replay_directory.py
• Extension examples: $ISAAC_SIM_SRC/extensions/isaacsim.replicator.mobility_gen.examples/
• Data dir: $MOBILITY_GEN_DATA (env var). Default to a workspace-local path such as $WORKSPACE_DIR/MobilityGenData or $HOME/MobilityGenData.
  • recordings/ — timestamped trajectory dirs
  • replays/ — rendered output
  • maps/ — occupancy map YAML + PNG files

Extension Loading (Critical)

Extensions are not auto-loaded. Always pass --enable flags when running python.sh:

"$ISAAC/python.sh" my_script.py --enable isaacsim.replicator.mobility_gen.examples

"$ISAAC/python.sh" my_script.py \
  --enable isaacsim.asset.gen.omap \
  --enable isaacsim.replicator.mobility_gen.examples

Enabling isaacsim.replicator.mobility_gen.examples auto-loads isaacsim.replicator.experimental.mobility_gen as a dependency. All extension-dependent imports must come AFTER SimulationApp(...) is initialized.

Import public names from the package root:

from isaacsim.replicator.experimental.mobility_gen import (
    ROBOTS, SCENARIOS, OccupancyMap, RecordingSession, load_scenario,
)

OccupancyMapDataValue is not re-exported; import it from ...mobility_gen.impl.occupancy_map.

Phase 1: Automated Trajectory Recording (Headless)

KeyboardTeleoperationScenario and GamepadTeleoperationScenario require an interactive UI. For headless batch recording use RandomPathFollowingScenario or RandomAccelerationScenario.

API note (Kit 110): MobilityGen no longer uses the legacy World flow. get_world() / new_world() are gone, and so is impl.utils.global_utils. Recording is driven by RecordingSession plus the isaacsim.core.simulation_manager.SimulationManager lifecycle, with isaacsim.core.experimental.utils.stage for stage I/O (note open_stage() returns a (bool, stage) tuple, and save_stage() takes only a path).

Migration: for the full omni.isaac.* → isaacsim.* mapping when porting scripts off the legacy World flow, see Renaming Extensions.

record_trajectories(scene_usd, omap_yaml, robot_type, scenario, num_episodes, max_steps, data_dir) — headless SimulationApp loop that builds a robot and scenario and records each episode to $MOBILITY_GEN_DATA/recordings/.

RecordingSession call order — the session owns the ground plane, robot spawn, Config and writer, so scripts do not construct a MobilityGenWriter themselves:

session = RecordingSession()
session.build(robot_cls, scenario_cls, occupancy_map,
              scene_usd=..., cached_stage_path=..., recordings_dir=...)
omni.timeline.get_timeline_interface().play()   # initialize() expects a playing app
simulation_app.update()
session.initialize()
session.reset()
session.enable_recording()
while ...:
    SimulationManager.step(steps=1)   # initialize_physics() does not start the
    simulation_app.update()           # timeline, so update() alone won't tick physics
    if not session.step(robot_cls.physics_dt):
        break

See scripts/record_trajectories.py.

Phase 2: Replay & Render

Replay all recordings in $MOBILITY_GEN_DATA/recordings/ and write sensor data to replays/.

: "${MOBILITY_GEN_DATA:=${WORKSPACE_DIR:-$HOME}/MobilityGenData}"
ISAAC="$ISAAC_SIM_DIR"
SRC="$ISAAC_SIM_DIR/source"

CUDA_VISIBLE_DEVICES=0 DISPLAY=:99 nohup \
  "$ISAAC/python.sh" \
  "$SRC/standalone_examples/replicator/mobility_gen/replay_directory.py" \
  --input  "$MOBILITY_GEN_DATA/recordings" \
  --output "$MOBILITY_GEN_DATA/replays" \
  --render_interval 40 \
  --rgb_enabled True \
  --depth_enabled True \
  --segmentation_enabled True \
  --normals_enabled False \
  --render_rt_subframes 1 \
  --enable isaacsim.replicator.mobility_gen.examples \
  > /tmp/mobility_gen_replay.log 2>&1 &

--render_interval 40 = 1 frame per 40 physics steps (~5 Hz at 200 Hz physics). Increase --render_rt_subframes for better quality at the cost of speed.

Replay Output Structure

replays/<recording_name>/
  config.json
  stage.usd
  occupancy_map/map.yaml, map.png
  state/
    common/<step>.npy          # robot pose, joint positions, velocities
    rgb/<camera_name>/<step>.jpg
    segmentation/<camera_name>/<step>.png
    depth/<camera_name>/<step>.png   # 16-bit inverse depth
    normals/<camera_name>/<step>.npy

Available Robots

The four *MultiSensorRobot variants subclass MobilityGenMultiSensorRobot and declare their cameras in a YAML sensor-rig config instead of the front_camera_* class attributes used by the single-camera robots above.

Available Scenarios

RandomPathFollowingScenario plans an A* path from the robot's current position to a random free-space goal and follows it with proportional steering. Episode ends when goal is reached or robot collides.

Add a Custom Robot

Two base classes exist depending on robot type. Both handle build() and write_action() — set class-level attributes only.

Wheeled (differential drive)

Subclass WheeledMobilityGenRobot. No need to override build() or write_action():

MyRobot(WheeledMobilityGenRobot) — example class showing all required class-level attributes (camera offsets, occupancy-map params, velocity ranges, wheel geometry) with no method overrides needed.

See scripts/wheeled_robot_subclass.py.

Reference implementations in isaacsim.replicator.mobility_gen.examples.robots:

• JetbotRobot: NVIDIA Jetbot, wheel_base=0.1125, wheel_radius=0.03, chassis_subpath="chassis"
• CarterRobot: Nova Carter, wheel_base=0.413, wheel_radius=0.14, chassis_subpath="chassis_link"

Holonomic (e.g. Kaya 3-wheel)

Override build() to use a different controller and write_action() to remap the 2D action:

KayaRobot(WheeledMobilityGenRobot) — overrides build() to configure a HolonomicController from HolonomicRobotUsdSetup, and write_action() to map [lin, ang] to [forward, lateral=0, yaw].

See scripts/holonomic_robot_subclass.py.

Policy-based (legged robots)

Subclass PolicyMobilityGenRobot and implement build_policy(). write_action() converts the 2D action [lin_vel, ang_vel] into the 3D command [x, 0, yaw] automatically. The policy spec selects the engine-specific robot USD.

Class attributes are the same occupancy_map_*, random_action_*, and path_following_* set as the wheeled robot, plus articulation_path and controller_z_offset.

@ROBOTS.register()
class MyLeggedRobot(PolicyMobilityGenRobot):
    physics_dt: float = 0.005
    z_offset: float = 1.05
    articulation_path = "pelvis"
    controller_z_offset: float = 1.05

    @classmethod
    def build_policy(cls, prim_path: str) -> RobotPolicyRunner:
        return RobotPolicyRunner(
            get_h1_spec(),
            prim_path=prim_path,
            position=np.array([0.0, 0.0, cls.controller_z_offset]),
        )

Reference implementations: H1Robot (articulation_path="pelvis") and SpotRobot (articulation_path="/") in the same module.

Replay with a custom robot

replay_directory.py calls load_scenario() which does ROBOTS.get(config.robot_type). If the robot isn't in the built-in extension, this raises KeyError. You cannot pass --enable to load an ad-hoc Python file — either create a proper Isaac extension, or copy the replay loop into your own script and register the robot class before calling load_scenario():

replay_with_custom_robot(input_dir, custom_robot_class) — register a custom robot class at runtime, then call load_scenario() for each recording directory.

See scripts/replay_custom_robot.py.

Config / Data Format

config.json per recording:

{
  "scenario_type": "RandomPathFollowingScenario",
  "robot_type": "CarterRobot",
  "scene_usd": "/path/to/warehouse.usd"
}

state/common/<step>.npy is a numpy dict: position, orientation, joint_positions, joint_velocities, linear_velocity, angular_velocity.

Common Pitfalls

• ModuleNotFoundError: No module named 'isaacsim.replicator.experimental.mobility_gen': Extensions aren't auto-loaded. Pass --enable isaacsim.replicator.mobility_gen.examples to python.sh. All extension imports must come AFTER SimulationApp(...).
• KeyError: 'CarterRobot' from ROBOTS.get(...) despite a clean import: ROBOTS came from a different registry than the one the examples extension populates. Import it from isaacsim.replicator.experimental.mobility_gen.
• ImportError: ...impl.utils.global_utils or get_world / new_world / join_sdf_paths undefined: removed with the legacy World flow. Use RecordingSession + SimulationManager, and isaacsim.core.experimental.utils.prim.join_prim_paths.
• Recording runs but every episode has 0 steps: session.step() was called without advancing physics. SimulationManager.initialize_physics() does not start the Kit timeline, so simulation_app.update() alone does not tick physics — call SimulationManager.step(steps=1) each iteration.
• Replay KeyError: 'MyRobot': replay_directory.py only knows built-in robots. Write a wrapper script that registers your robot class before calling load_scenario().
• Custom robot produces no images during replay: Missing front_camera_* attributes, or build() passes front_camera=None. Add the attributes and call cls.build_front_camera(prim_path) in build().
• AttributeError: 'MyRobot' has no attribute 'chase_camera_base_path': chase_camera_base_path, chase_camera_x_offset, chase_camera_z_offset, chase_camera_tilt_angle are required by load_scenario() even for headless recording.
• Built-in replay fails to find robot class: Pass --enable isaacsim.replicator.mobility_gen.examples so the examples extension registers its robots/scenarios before load_scenario() runs.
• physics_dt mismatch: Recording stores the physics timestep in config.json; replay uses the same robot_type.physics_dt. Do not change robot params between record and replay.
• Occupancy map scale: MobilityGen consumes the OccupancyMap produced by occupancy-map. Ensure map.yaml origin and resolution match the USD world coordinates.
• Headless GPU: Set CUDA_VISIBLE_DEVICES=0 DISPLAY=:99 to avoid GPU contention with vLLM on GPUs 1-3.