본문으로 건너뛰기
cccskills
무료GitHub에 공개

isaac-humanoid-cloud-control

Run a Unitree G1 navigation demo in Isaac Sim through cloud control. The skill coordinates browser streaming, route checks, operator approvals, mission submission, and monitoring, with a guarded process for releasing the simulated support gantry.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

원문원문 설명 보기

Run streamed Unitree G1 cloud-navigation demos in Isaac Sim with guarded gantry release. Use for demo, destination, Mission Control, or cleanup requests.

연동 도구Docker

실행 점수 미검증

설치 방법 보기

이럴 때 유용합니다

  • Run a cloud-controlled humanoid navigation demo
  • Move toward a selected object in the simulated scene
  • Start the demo stack for inspection
  • Clean up resources owned by the demo

포함된 파일(50)

  • SKILL.md13.2 KB
  • assets/browser-viewer.compose.yaml921 B
  • assets/carter_warehouse_navigation.png4.5 KB
  • assets/carter_warehouse_navigation.yaml280 B
  • BENCHMARK.md10.5 KB
  • config/skillspector-baseline.yaml2.3 KB
  • evals/evals.json11.7 KB
  • evals/evals.json.license137 B
  • extensions/humanoid_cloud_control.autoplay/config/extension.toml459 B
  • extensions/humanoid_cloud_control.autoplay/humanoid_cloud_control/autoplay/__init__.py171 B
  • extensions/humanoid_cloud_control.autoplay/humanoid_cloud_control/autoplay/extension.py2.1 KB
  • references/bootstrap.md5.4 KB
  • references/bring-up-cloud-stack/README.md1.2 KB
  • references/bring-up-cloud-stack/scripts/run.py1.6 KB
  • references/change-fleet-composition/README.md1.1 KB
  • references/change-fleet-composition/scripts/run.py1.6 KB
  • references/change-map/README.md1.0 KB
  • references/change-map/scripts/run.py1.6 KB
  • references/cloud-workspace.md4.7 KB
  • references/dependencies.md6.5 KB
  • references/isaac-ros-activate/README.md1.9 KB
  • references/isaac-ros-activate/scripts/run.py1.6 KB
  • references/isaac-ros-setup.md7.7 KB
  • references/isaac-sim-installation/README.md720 B
  • references/isaac-sim-remote/README.md1.1 KB
  • references/isaac-sim-remote/scripts/run.py1.6 KB
  • references/ordered-startup.md19.9 KB
  • references/release-4.6.md6.7 KB
  • references/submit-and-monitor-mission/README.md1.6 KB
  • references/teardown.md1.3 KB
  • references/troubleshooting.md1.5 KB
  • scripts/check_runtime.py5.0 KB
  • scripts/doctor.py17.2 KB
  • scripts/guarded_gantry_release.py12.8 KB
  • scripts/http_transport.py2.2 KB
  • scripts/play_viewer.py4.2 KB
  • scripts/preflight.py18.3 KB
  • scripts/prepare_cli_repository.py2.9 KB
  • scripts/scene_inventory.py6.7 KB
  • scripts/simulation_identity.py968 B
  • scripts/submit_navigation_mission.py10.4 KB
  • scripts/suggest_navigation_missions.py10.3 KB
  • shared/dependency_manifest.py6.5 KB
  • shared/simulation_gate.py2.2 KB
  • skill-card.md4.8 KB
  • skill.oms.sig15.5 KB
  • tests/test_dependencies.py6.9 KB
  • tests/test_io_boundaries.py5.9 KB
  • tests/test_security_gates.py5.7 KB
  • upstream-versions.lock.json4.3 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Isaac Humanoid Cloud Control

Purpose

Run the public NVIDIA Isaac ROS release-4.6 Unitree G1 cloud-control workflow as an ordered integration. Default to the Isaac Sim full streaming app running headless, an interactive browser WebRTC view, and a navigation mission. Do not silently fall back to a local Isaac Sim window. Read the release-4.6 runbook before executing the workflow.

Inputs

Required inputs are the request mode (bring-up, end-to-end demo, destination, or cleanup) and, before mission submission, an operator-supplied destination and movement approval. A selected suggestion supplies both for simulation. Remote missions also require confirmation that the rendered viewport is visible.

Optional or discoverable inputs are robot identity (g1_0 by default), hosts, spawn pose, map, and installation paths. Resolve them from active demo state, explicit operator choices, live observations, discovered installations, then packaged defaults. Never let a default override an explicit safety boundary.

Prerequisites

Require an x86_64 NVIDIA GPU with NVENC, Isaac Sim 6.0, the canonical Docker-mode Isaac ROS workspace, NVIDIA's official web viewer, and the release-4.6 cloud images. Use the bootstrap gate on a fresh host. If the CLI or its Docker environment is missing, follow CLI installation and Docker setup, then rerun preflight. Missing CLI installation is a recovery step, not a terminal blocker.

Preserve the target boundary

  • Require an NVIDIA GPU with NVENC for the default livestream. Preflight TCP 49100 for WebRTC signaling, UDP 47998 for media, and TCP 8210 for the browser viewer. Keep the endpoints on a private or otherwise trusted network; they provide no authentication or encryption by themselves.
  • Use the owning cloud skill's broker authentication configuration and supply matching credentials to the G1 client. Keep broker ports on the host or a trusted private network; authentication alone does not provide TLS.
  • Enable Isaac Sim's Python server only for live scene inventory. Keep it on loopback at 127.0.0.1:8226; it executes Python and must never be published, proxied, or exposed to the browser network.
  • Use an x86_64 NVIDIA GPU host and Isaac Sim 6.0 for the sample simulation.
  • Support navigation through cloud control in Isaac Sim.
  • For Isaac Sim-only navigation, use enable_camera_localization:=false and use_static_tf:=true. Do not launch, configure, validate, or require Visual SLAM, visual global localization, RealSense, D435, or camera-localization topics; they are outside this simulated-demo path.
  • Do not submit humanoid_manipulation objectives to Isaac Sim. GR00T cloud manipulation is real-hardware-only in this release.
  • Treat real G1 operation as a separate safety-gated path. Require the Unitree network setup, D435 localization, CUDA MPS, real-time scheduling, and the lower-body-then-upper-body controller enable sequence from the upstream GR00T guide. Never infer permission to move real hardware.
  • Submit a navigation mission only after the operator has supplied the target and explicitly approved the previewed mission payload. Verify occupancy, robot-footprint clearance, and Nav2 reachability before submission. A target selection alone is insufficient; changed routes require a new approved preview.
  • Before invoking the guarded helper, obtain explicit operator approval for that virtual gantry release attempt. A broad demo request is insufficient. State the simulation target and action; ask “Release the Isaac Sim virtual gantry now?” unless the operator already explicitly approved this specific attempt. Pass the live simulation identity interlock and unsupported stability check before offering destinations. A retry needs separate explicit approval after reporting the failed attempt and confirmed gantry re-engagement. Never apply this automation to hardware. Honor an operator's request to use the viewport G action instead.
  • Keep one robot identity everywhere. Default to g1_0; apply the same value to Mission Control, serial_number, MQTT topics, mission payloads, and lifecycle verification.
  • Use one dedicated, initially empty ROS_DOMAIN_ID for Isaac Sim and every ROS shell in a demo. Set localhost-only discovery for all demo processes as specified in the ordered startup. Confirm that the clock and robot-description ROS topics are absent on that domain before startup. This prevents transient-local samples from a previous crashed run from racing the new controller manager.
  • The bundled occupancy-grid PNG supplies obstacle and free-space pixels for the navigation map; its YAML metadata defines the map scale and origin. The cloud workspace procedure copies this pair into the runtime workspace. These are map data, not executable instructions.
  • Keep one occupancy map everywhere. The robot launch and Mission Control must consume the same PNG and YAML pair.
  • Ensure the map YAML and referenced image are readable at the same absolute path inside the activated Isaac ROS environment; host-only paths are insufficient.
  • Create a dedicated runtime MC_WORKSPACE using the cloud workspace procedure. Generate map and fleet configuration through the owning adapters. Resolve Compose, image references, and supporting resources from the pinned upstream skills; never modify their checkouts or store generated configuration in this package.

Resolve upstream scripts

Set SKILL_DIR to the absolute directory containing this loaded skill manifest, using the skill loader's location, and preserve it in every shell invocation. All bundled paths are relative to it; never derive them from the working directory. Read the dependency workflow before calling another skill's scripts. It follows the Mission Control showcase composition model: pinned external checkouts, a persisted manifest, and parent-owned adapters. Read each pinned upstream instruction manifest from the resolved dependency record and follow its gates. The exact source paths and commits are disclosed in the dependency lock. Adapter contracts: cloud, map, fleet, scene, and installation handoff, and ROS activation, and mission verification.

Resolve dependencies before host preflight; resolution does not launch services:

python3 "${SKILL_DIR}/scripts/doctor.py" dependencies

Normal invocation never downloads missing skills. Explicit dependency preparation uses dependencies --prepare only after authorization as described in the reference. Use the linked adapters instead of guessed upstream script paths.

Compose the owning skills

Follow owning-skill responsibilities for each handoff; the upstream skill remains authoritative.

Available Scripts

ScriptPurposeMain arguments
scripts/doctor.pyResolve pinned upstream scripts and write their manifestdependencies; optional --prepare, --check-only, --json
scripts/prepare_cli_repository.pyStage and verify the public APT key and fixed repository source without rootNo arguments
scripts/preflight.pyDiscover and validate host, cloud, streaming, and scene-server readiness--map; optional installation paths and --check-* gates
scripts/guarded_gantry_release.pySet AGILE, verify stability, release the simulated gantry, and verify unsupported balance--allow-simulation-gantry-release; stability limits and durations
scripts/check_runtime.pyVerify ROS clock, sensors, TF, and Nav2 lifecycle--base-frame, --timeout
scripts/scene_inventory.pyInventory live-stage objects without modifying USDinjected output_path, max_objects
scripts/suggest_navigation_missions.pyProduce occupancy- and footprint-safe object approach poses--inventory, --map, optional robot pose
scripts/submit_navigation_mission.pyPreview or submit a validated Mission Control route--host, --robot, waypoints, --validated-free-space, --submit
scripts/play_viewer.pyFocus the browser viewer and send Play only as a documented fallback--url, browser and debug options

Use run_script() for these helpers when the runtime exposes it; otherwise run the bundled helper by its resolved absolute path without reconstructing it.

Fresh-machine bootstrap gate

For a fresh host, read the bootstrap gate, run its read-only discovery first, and resolve failures in the documented ownership order. Continue only after preflight has no blocking failures. Report the exact failed observation rather than claiming a live view is available.

Instructions

Read and follow the ordered startup and mission workflow in full before executing. It contains the exact commands and phase gates:

  1. Discover host paths and inspect existing state read-only; resolve upstream dependencies before host preflight.
  2. Prepare the release package and Deploy extension, and preview and apply requested map and fleet changes through the owning adapters.
  3. Start cloud services through the cloud adapter, with the runtime workspace and no generic robot simulator.
  4. Prepare the official browser viewer and publish its reachable URL.
  5. Launch rendered headless Isaac Sim paused, stage the activated G1 ROS launch, then start the timeline and ROS stack in the required order.
  6. Verify AGILE and run the guarded simulation-only gantry release. Require unsupported balance before inventorying or offering destinations.
  7. Verify runtime and cloud readiness; inventory the live scene through the remote adapter and validate navigation suggestions with occupancy and Nav2.
  8. After destination approval, validate and submit once, keep the live view visible, and apply submit-and-monitor-mission until completion.

The detailed workflow is mandatory; this index does not replace its commands, startup timing, retry limits, or authorization boundaries.

Troubleshooting

Apply the readiness and troubleshooting checklist before submission and when diagnosing a failed gate.

Examples

  • “Run an end-to-end humanoid cloud-control Isaac Sim demo” starts setup; obtain explicit virtual gantry release approval before running the helper.
  • “Destination 1: forklift” after live suggestions selects the target. Validate the route, preview the mission payload, and obtain submission approval.
  • “Bring up the stack for inspection” does not authorize gantry release or a mission.
  • “Clean up the stack” authorizes teardown of demo-owned resources only.
  • “Use GR00T to manipulate a box in Isaac Sim” is outside this release; offer supported navigation without inventing manipulation support.

Output Format

Report these fields in order when applicable:

  1. State — current phase or first failed gate with measured evidence.
  2. Viewer — clickable browser URL while live and visibility status.
  3. Mission — ID, terminal state and timestamp, initial, moving, and final poses, Dispatch and Nav2 evidence, and final robot health.
  4. Cleanup — demo-owned resources removed, reusable resources preserved, and confirmation that the previous viewer URL is unavailable.

Limitations

Support cloud-controlled navigation only in Isaac Sim release 4.6. Do not claim simulated GR00T manipulation, operate real hardware, expose unauthenticated demo endpoints publicly, replace missing owning skills, or delete shared data. Treat camera localization and real-G1 controller sequencing as separate workflows.

Stop behavior

Keep the completed demo running by default. After a successful mission, leave the robot launch, headless Isaac Sim, browser viewer, and cloud services alive so the user can inspect the result or request another mission. Keep publishing the clickable live-stream URL in the handoff and offer an explicit stack cleanup. Do not infer cleanup from mission completion, live-view confirmation, the end of the current turn, or a request that names only one finished mission.

Clean up only on an explicit teardown request. Then read and apply the teardown procedure. Remove only resources owned by the demo run, preserve shared data and the reusable Isaac ROS container, verify the demo endpoints are closed, and invalidate the prior viewer URL.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Move pandas-style table processing to NVIDIA GPUs using cuDF. Choose minimal-change acceleration or explicit GPU code, handle joins and aggregations, compare results with pandas, and use dask-cuDF when data exceeds GPU memory.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

  • Accelerate existing pandas processing with minimal changes
  • Move joins and aggregations to GPU code
  • Compare missing values and ordering with pandas
NVIDIA/skills3.7천2026년 10월 24일 업데이트

Customize NVIDIA Nemotron Voice Agent’s Generic Pipecat example for appointment scheduling, five-field patient intake, or a custom healthcare conversation. Build prompts and tool handlers without a separate backend, then validate conversation behavior, services, and voice interaction.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

  • Build an appointment-scheduling voice agent
  • Collect five patient-intake fields through conversation
  • Design a custom healthcare conversation with tool calls
NVIDIA/skills3.7천2026년 10월 24일 업데이트

Calibrate live cameras by recording their RTSP video streams through VIOS and processing the clips with AutoMagicCalib. The workflow imports settings and alignment files, checks readiness, tracks calibration progress, and retrieves results.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

  • Calibrate cameras from live RTSP streams
  • Import settings and alignment files for a camera scene
  • Retrieve evaluation statistics using ground truth data
NVIDIA/skills3.7천2026년 10월 24일 업데이트

Test a running AMC service with the bundled four-camera warehouse dataset. Run camera calibration from upload to completion, then check distance and image projection errors against ground-truth data to verify the installation.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

  • Verify a newly installed AMC service
  • Check the sample's calibration errors against ground truth
  • Walk through the sample workflow in Swagger UI
NVIDIA/skills3.7천2026년 10월 24일 업데이트

Calibrate multiple cameras from recorded MP4 files through the AutoMagicCalib REST API. The skill handles project creation, uploads, input verification, progress checks, and evaluation when reference data is supplied, with optional VGGT refinement.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

  • Calibrating cameras from synchronized MP4 recordings
  • Uploading local settings and alignment files
  • Evaluating calibration against ground-truth data
NVIDIA/skills3.7천2026년 10월 24일 업데이트

Set up AutoMagicCalib from NGC release containers using Docker Compose. The skill handles registry authentication, port and path configuration, optional VGGT model download, service startup, and readiness checks for the backend and web interface.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

  • Launching the AutoMagicCalib backend and web interface
  • Configuring service ports and network access
  • Adding the optional VGGT model during setup
NVIDIA/skills3.7천2026년 10월 24일 업데이트

NVIDIA의 스킬 모두 보기

이 스킬의 문제 신고