isaac-mission-control-showcase

Mission Control Showcase

Use the bundled runner for the default small-warehouse Nova Carter showcase. It owns the integration and coordinates five upstream skills that it resolves at runtime rather than vendoring. Resolve those dependencies once with python3 "$SKILL_DIR/scripts/doctor.py" dependencies --prepare before the first run.

Locate bundled resources

Before running any command, set SKILL_DIR to the absolute path of the directory containing this loaded SKILL.md, using the skill location supplied by the skill loader. Preserve that value in each shell invocation. Do not derive it from the current working directory or assume a source repository layout or a fixed installation directory.

All bundled paths in this document and its references, including scripts/, references/, shared/, assets/, config/, and upstream-versions.lock.json, are relative to SKILL_DIR. Read them beneath that directory and prefix bundled script paths with "$SKILL_DIR/" when following command examples in the references. Runtime output paths and upstream skill paths reported by the dependency manifest retain their own locations. Commands below can run from any working directory.

Runtime selection

Always launch Isaac Sim from a local installation through this showcase runner, with its GUI window on the host display. Watch the robot in that window. Drive every stage operation through the isaac-sim-remote Python server on ISAAC_PYTHON_PORT. There is no WebRTC, browser viewer, or containerised Isaac Sim in this workflow.

Before preflight, inspect existing state read-only. Determine the installation path and version, whether isaac-sim.sh and the bundled ROS 2 bridge libraries are present, whether a display is available, and whether an Isaac Sim is already running. Do not launch or import Isaac Sim, start or stop a container, run an installer, or alter files.

Treat the installed version, warehouse URI, Nova Carter asset and sensor graphs, and readiness probes as one compatibility set. Mere presence of Isaac Sim is not proof of compatibility.

Apply this selection policy:

• If a 6.1.0 installation is present at ISAAC_SIM_DIR, report the detected version and continue without an unnecessary prompt.

• If a different version is detected and the launcher, ROS 2 bridge, and a verified warehouse URI are all present, use it automatically — no prompt — and report the choice:

  Isaac Sim <version> detected at <path> and is usable (launcher <ready/status>, ROS 2 bridge <ready/status>). Using this installation.

  Attempt the normal preflight and readiness contract; stop and report a mismatch rather than modifying that installation.

• Prompt before preflight only when the version cannot be determined, or none of launcher, ROS 2 bridge, or a verified warehouse URI can be confirmed — this is the sole runtime-selection case that requires user input, because no automatic choice can be verified safe:

  Isaac Sim <version> was detected at <path>. Launcher: <ready/status>. ROS 2 bridge: <ready/status>. Choose whether to use the detected version or select or install another version.

• Docker is still required for the Mission Control cloud stack and Nova Carter SIL, but never for Isaac Sim itself.

• If nothing is installed, route to the resolved isaac-sim-installation skill rather than guessing a path. Resolve it through python3 "$SKILL_DIR/scripts/doctor.py" dependencies; never search for it and never vendor a copy.

Never upgrade, downgrade, overwrite, delete, repair, relabel, refresh, or otherwise modify an existing Isaac Sim installation as showcase recovery. Obtain explicit user direction before stopping an already-running Isaac Sim.

If the user asks to install, upgrade, downgrade, or otherwise obtain a different Isaac Sim version, stop the showcase workflow and hand off to the resolved isaac-sim-installation skill. Read its runtime SKILL.md at the path the dependency manifest reports, and follow its own gates without bypassing any of them. Do not reproduce or restate its installation logic here. The runner requires a local standalone installation, so preselect that method. Stop after the installed path is reported; do not launch Isaac Sim. Resume runtime selection only after that handoff has completed and the user asks to continue. When that skill runs a compatibility check, prefer a generous --timeout-seconds such as 3600; its 600-second default often expires on a first run that is still populating shader and asset caches.

Default workflow

1. Check whether the user already specified a different map, robot platform, simulator, or mission objective in the conversation; if not, use the canonical default without asking.

2. Apply Runtime selection without changing any detected installation. Resolve any user choice or installation handoff, then run the read-only preflight:

   bash "$SKILL_DIR/scripts/run.sh" --preflight

3. Ask the user to stop or relocate every reported running container that is not explicitly permitted. Allow unrelated workloads by exact container name with a repeated --ignore-container NAME; ignored containers still undergo the normal port-conflict checks. Do not stop an existing Isaac Sim automatically. Require a free ISAAC_PYTHON_PORT.

4. Start the stack and leave the robot idle:

   bash "$SKILL_DIR/scripts/run.sh"

5. Or run the deterministic closed route and wait for completion:

   bash "$SKILL_DIR/scripts/run.sh" --demo

6. Watch the robot in the Isaac Sim window that the runner opens. The runner loads the stage before Carter starts, so initial robot motion remains visible.

7. Stop only the recorded showcase resources:

   bash "$SKILL_DIR/scripts/run.sh" --stop

Every run writes run-manifest.json into its work directory before anything starts and updates it at each transition, so an interrupted run can still be described and stopped precisely. run-result.json is the machine-readable acceptance artifact. Inspect a run, including one that was interrupted, with:

python3 "$SKILL_DIR/scripts/showcase.py" \
  run-status --work-dir <dir>

A stop targets only the containers that run recorded, is idempotent, and preserves logs and evidence. Motion evidence is persisted when observed, so a later stationary sample never overrides a confirmed observation. Acceptance stays strict regardless: COMPLETED mission, robot online, healthy and IDLE, motion confirmed, no OOM, required processes healthy.

See references/troubleshooting.md for the lifecycle states, recovery procedure, motion verdicts, Isaac Sim exit handling, host memory policy, and the ROS 2 bridge executable check.

The runner creates a fresh directory under ${TMPDIR:-/tmp}, prints its path, and writes all generated Mission Control configuration, maps, Isaac caches, logs, and state there. It must not create runtime files in the caller's working directory. If startup or demo execution fails after the Nova Carter container is created, the runner saves its complete timestamped ROS 2 launch output to $SHOWCASE_WORK_DIR/logs/nova-carter-ros2.log. It refreshes the same log after stopping Nova Carter with --stop.

Composition model

Two layers.

Bundled and parent-owned — everything in this package: references/, shared/, scripts/, assets/, config/. The references/ tree holds integration adapters and routing contracts written for this showcase. They are not copies of upstream skills.

Runtime-resolved — the five skills the showcase orchestrates. They are owned by their own repositories, read and invoked from pinned checkouts outside this package, and never copied in.

In the table below, adapter paths are relative to the corresponding reference directory.

upstream-versions.lock.json is the single dependency declaration: the two public GitHub repositories, their pinned refs and immutable commits, and each skill's required entrypoints and resources.

Resolve once per host, then run:

python3 "$SKILL_DIR/scripts/doctor.py" dependencies --prepare \
  --env-file "$HOME/.mission-control-showcase/state/showcase-deps.env"

scripts/doctor.py runs both phases: dependencies resolves the upstream skills and writes the manifest, host checks this machine. The dependency phase must succeed first, because the host phase is handed the entrypoints it resolved.

Read references/workflow.md for the stage router, and each reference's README.md before using that stage.

Using an upstream skill

The reference tells you where the upstream is; the upstream tells you how it behaves.

1. Read its runtime SKILL.md, at the path the manifest records as skills.<name>.skill_md. Nothing in this package restates it.
2. Follow its own gates, including confirmation, licensing, and execution-approval gates.
3. Invoke it through the reference adapter, which runs it from its own skill directory so relative resources resolve there.
4. Never reimplement it, never vendor it, and never substitute another copy found on the machine.

Do not invoke bring-up-cloud-stack with --with-sim; that starts the generic mission-simulator, not Nova Carter SIL. The showcase owns Nova SIL, Isaac Sim launch and shutdown, scene restore, cross-domain readiness, mission submission, and completion evidence.

Missing dependencies

The runner performs a read-only freshness check and the adapters consume the persisted manifest. Normal execution never clones, fetches, pulls, installs, or updates anything.

With MISSION_CONTROL_SHOWCASE_REQUIRE_PREFLIGHT=1, a reference that cannot load a ready manifest blocks rather than falling back to discovery. An adapter exits 3 with BLOCKED [<component>]: … naming the skill, what was expected, and the preflight command. Never work around a blocker by vendoring a skill.

Canonical default

The default is:

• Isaac Sim: a local installation at ISAAC_SIM_DIR (default ~/isaacsim), launched with its GUI window. Isaac Sim 6.1.0 is the last-known-good runtime. This is a default, not a compatibility ceiling.
• Warehouse: https://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/Isaac/6.1/Isaac/Environments/Simple_Warehouse/warehouse.usd.
• Nova Carter SIL: nvcr.io/nvidia/isaac/nova_carter_sil:release-3.2.
• NVIDIA driver R590 or newer.
• A 64 GB-class host with at least 48 GB available before startup and 4 GB free swap.
• Nova Carter SIL is limited to 20 GB RAM plus up to 4 GB swap.
• Robot identity: carter01.
• Isaac Sim Python server on port 8226.
• Websocket MQTT on port 9001, matching the packaged Mission Control stack.
• ROS domain 77 for both Isaac Sim and Carter SIL, with normal host transport. The isolated domain prevents unrelated domain-0 DDS participants from exhausting Carter SIL memory while preserving Isaac-to-SIL discovery.

The exact warehouse URI is passed into Isaac Sim and verified after stage load. Never discover the warehouse through get_assets_root_path, a local cache, a source checkout, or a filesystem search.

The runner configures Mission Control and starts the cloud stack, then launches Isaac Sim and restores the stage while the cloud initializes. Carter SIL starts last, so the Isaac Sim window shows the robot's initial motion.

For a user-selected noncanonical Isaac version that is already installed:

1. Require a compatible installation and a canonical asset URI.

2. Pass both explicitly:

   ISAAC_SIM_DIR=<resolved-install-path> \
   WAREHOUSE_USD_URI=<resolved-uri> \
     bash "$SKILL_DIR/scripts/run.sh"

3. Do not infer a future URI by substituting a version number. A local installation carries no inspectable version tag, so the caller owns the asset contract.

4. If the requested version is not already installed, stop and hand off to isaac-sim-installation; do not install it inside this workflow.

Robot identity

carter01 is only the default. Set another identity with either:

bash "$SKILL_DIR/scripts/run.sh" \
  --robot-name my_robot --demo

or ROBOT_NAME=my_robot.

The runner applies the same value to Mission Control fleet configuration, Mission Client serial_number, MQTT bridge client names, readiness polling, and mission submission. Do not change only the mission payload.

Deterministic route

The default replay is a closed route:

[
  {"x": -5.1, "y": 1.4},
  {"x": -2.625, "y": 1.2},
  {"x": -0.2, "y": -2.075},
  {"x": -2.625, "y": -5.35},
  {"x": -5.1, "y": 1.4}
]

Submit it as a route-only mission with timeout 900 and solver NVIDIA_CUOPT. Do not add start/end locations or iterations.

For custom coordinates, first bring up the stack without --demo, snap the requested points through WPG nearest_nodes, validate with visualize_route, and submit only the resulting routable points.

Readiness contract

The runner must not submit a mission until all of these are observed:

• the canonical stage is active and the Isaac timeline is playing;
• ROS /clock advances;
• a live /scan LaserScan uses frame front_2d_lidar and has a timestamp within one second of simulation time;
• TF resolves the LiDAR frame into map at that exact sensor timestamp;
• map -> base_link resolves;
• the robot pose lies inside the active global costmap;
• Nav2 planner, controller, and navigator lifecycle nodes are active;
• Carter SIL has recorded no cgroup OOM event and its required map, localization, planning, control, navigation, and mission-client processes remain alive;
• Mission Control reports the configured robot online, IDLE, error-free, position-initialized, and near the reset pose.

Treat an initial exact-time LiDAR TF miss as a bounded startup condition. Continue consuming newer scan samples until one resolves into map, or fail at the readiness deadline with the latest measured timestamp. Never replace this with a latest-TF lookup.

On failure, report the failed observation and measured values. Do not claim a clock, TF, costmap, or permission root cause without those measurements.

The bundled carter_warehouse_navigation.png and companion YAML are the single map resource. The runtime copies that pair into Mission Control's generated config and Carter SIL's read-only /maps mount. Mission Control uses the same PNG and metadata to generate the WPG resource under carter_warehouse_navigation.png. The runner verifies the source, Mission Control, and Carter hashes before startup, then verifies the active WPG map identity, Carter launch option, and container-mounted bytes. Never recursively change permissions on a source tree or apply a permission fix after services start.

Modes and overrides

• --preflight: read-only installation, display, Docker, GPU, dependency, port, image, and asset checks.
• no flag: clean startup and full readiness validation, leaving the robot idle.
• --demo / --replay: startup, readiness, circular mission, canonical mission completion, and pose-motion evidence.
• --stop: stop the last recorded stack and SIGTERM the Isaac Sim process, without deleting containers, images, volumes, caches, or logs.
• --dry-run: print resolved commands without changing runtime state.
• --work-dir PATH: use a caller-selected new or empty runtime directory.
• --no-pull: require required images to already be local.
• --ignore-container NAME: allow one already-running container by exact name during preflight and startup; repeat the option for each allowed container. Port conflicts remain failures.

Supported environment overrides are ISAAC_SIM_DIR, WAREHOUSE_USD_URI, NOVA_CARTER_IMAGE, NOVA_CARTER_MEMORY_LIMIT, NOVA_CARTER_MEMORY_SWAP_LIMIT, ROBOT_NAME, SHOWCASE_GPU_DEVICE, SHOWCASE_ROS_DOMAIN_ID, SHOWCASE_ROS_LOCALHOST_ONLY, ISAAC_PYTHON_PORT, and SHOWCASE_WORK_DIR. Keep the default isolated domain unless it conflicts with another local ROS deployment. Apply the same domain and localhost setting to Isaac Sim and Carter SIL. Do not enable localhost-only transport without verifying cross-process discovery between the host Isaac Sim bridge and host-network Carter SIL.

Acceptance

A successful --demo requires Mission Database to report the submitted mission COMPLETED, the same robot to return to IDLE without errors, and Mission Database poses to prove motion. Robot state, logs, a visible Isaac Sim window, or a single screenshot alone are not completion.

Do not run docker compose down -v, remove containers, prune Docker state, or delete the run directory as automatic recovery.