Skip to content
AgibotTechPublic

About

Simulation Platform from AgiBot

Resources

Stars

1.7k stars

Watchers

11 watching

Forks

Latest commit

ย 

History

212 Commits

Folders and files

NameName
Last commit message
Last commit date
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

Genie Sim

1. ๐ŸŒ Genie Sim

Genie Sim Benchmark runs with system Python 3.12 and Isaac Sim 6.2. Use geniesim docker or geniesim docker6.2 to select the dedicated image, geniesim4:latest, on the 4.0 runtime line. See release lines and runtime selection.

โš ๏ธ Deprecated (E.O.L.): Isaac Sim 4.5, standalone 5.1, and the 5.1 + 6.2 combo. Their code remains until retirement. Use geniesim docker / docker6.2 for current workflows.

Genie Sim is the simulation platform from AgiBot. It provides developers with a complete toolchain for environment reconstruction, scene generalization, data collection, and automated evaluation. Its core module, Genie Sim Benchmark is a standardized tool dedicated to establishing the most accurate and authoritative evaluation for embodied intelligence.

The platform integrates 3D reconstruction with visual generation to create a high-fidelity simulation environment. It pioneers LLM-driven technology to generate vast simulation scenes and evaluation configurations in minutes. The evaluation system covers 200+ tasks across 100,000+ scenarios to establish a comprehensive capability profile for models. Genie Sim also opens over 10,000 hours of synthetic data including real-world robot operation scenarios.

The platform will significantly accelerate model development, reduce reliance on physical hardware, and empower innovation in embodied intelligence. Simulation assets, dataset, and code are fully open source.

2. โœจ Features

  • ๐Ÿงž geniesim CLI: one command-line entry point for the entire platform โ€” docker lifecycle, ROS 2 workspace builds, health diagnosis (doctor), pure-Python wheel deploy, asset / scene bootstrap. Packaged as a standalone PEP 517 / PEP 621 wheel.
  • ๐Ÿค– Agent-ready SKILLs: packaged SKILL.md recipes for benchmarking, inference checks, scene generation, asset search, and teleop โ€” invokable directly by Claude Code / agentic coding agents, so the simulator can be driven without a human first reading the docs.
  • โšก Genie Sim RT Engine (realtime, interactive): geniesim_ros brings physics, render, and the robot into ROS 2 natively โ€” teleop, MoveIt, and ros2_control close the loop on one shared sim_time. Multiple physics backends (Isaac Sim PhysX, Newton-standalone) selectable per scene, with cloth + soft-body support on the Newton path.
  • ๐ŸŽจ High-Fidelity Sim-Ready Assets: 5,140 validated 3D assets covering five real-world operation fields: retail, industry, catering, home and office. ModelScope.
  • ๐Ÿ›ฐ๏ธ 3DGS-based Reconstruction Pipeline: Integrate 3DGS-based reconstruction process with visual generative model to synthesize realistic simulation environment with high-precision meshes. ModelScope.
  • ๐ŸŒ Genie Sim World: A spatial world model which generates a photorealistic 3D world from an equirectangular panorama. GitHub.
  • ๐Ÿ’ฌ LLM-Driven Scene Generation: Natural language-driven generation and generalization which instantly generates diverse simulation scenes through conversational interaction.
  • ๐Ÿ“ฆ Large-Scale Synthetic Dataset: Over 10,000 hours open-source synthetic data across 200+ loco-manipulation tasks with multi-sensor streams, alongside multi-dimensional variations.
  • ๐Ÿ› ๏ธ Synthetic Data Generation: Efficient toolkit for data collection with error-recovery mechanism, supporting both low-latency teleoperation and automated data programming. ModelScope.
  • ๐Ÿงช Robust and Diverse Benchmark: Provide 100,000+ simulation scenarios and use LLM to autonomously generate task instructions and evaluation configurations. Discrepancy between simulation and real-world test results is less than 10%.
  • ๐Ÿ” VLM-based Auto-Evaluation System: Full-spectrum evaluation criteria to provide model's capability profile covering manipulation skills, cognitive comprehension and task complexity.
  • ๐Ÿš€ Zero-Shot Sim-to-Real Transfer: Model trained with our synthetic data exhibits zero-shot sim-to-real transfer capability with superior task success rate compared to model trained with real data.

3. ๐Ÿš€ Quick start

3.1 ๐Ÿงž CLI & ๐Ÿค– SKILLs โ€” install, build the container, get a shell

flowchart LR
  pip(["๐Ÿ“ฆ pip install<br/>geniesim_cli<br/><i>(host)</i>"]) ==> docker(["๐Ÿณ geniesim<br/>docker up<br/><i>(host)</i>"]) ==> auto(["โš™๏ธ tier-1 peers<br/>auto-install<br/><i>(in-container)</i>"]) ==> shell(["๐Ÿ’ป geniesim<br/>docker into<br/><i>(host)</i>"])
  classDef step fill:#dbeafe,stroke:#1e40af,color:#1e3a8a,font-weight:bold
  class pip,docker,auto,shell step
  linkStyle 0,1,2 stroke:#1e40af,stroke-width:2px
Loading
# === on the host ===
pip install -e source/geniesim_cli/      # standalone CLI wheel
git submodule update --init -- source/geniesim_ros/src/ros_ws/src/genie_sim_robot_model
geniesim docker build                    # build the Isaac Sim 6.2 + Jazzy image
geniesim docker up                       # boot the container; the entrypoint
                                         # editable-installs every tier-1 from
                                         # the bind-mounted workspace
geniesim docker into                     # interactive shell inside the container

# === inside the container โ€” sanity-check the stack ===
geniesim status                          # per-distribution health probe
geniesim doctor                          # diagnose & repair (rosdep, env, โ€ฆ)
geniesim bootstrap                       # install or repair required Tier-1 peers

To skip geniesim docker build, follow the prebuilt-image instructions, then continue with geniesim docker up / into.

๐Ÿค– With an AI assistant โ€” just ask

The repo ships AGENTS.md + per-package skills/<name>/SKILL.md as machine-readable signposts. Claude Code, Cursor, etc. follow them on their own โ€” say "run the pnp benchmark against localhost:8999" or "launch the WBC scene with MoveIt" and the agent walks the trail.

๐Ÿ“– Without an AI assistant โ€” follow the signposts

AGENTS.md (root, boot sequence) โ†’ source/AGENTS.md (module map) โ†’ source/<pkg>/AGENTS.md (architecture + commands) โ†’ source/<pkg>/skills/<name>/SKILL.md (copy-paste recipes). Full CLI verb table: .agent/geniesim_cli.md.

Important

๐ŸŽจ Tier-1 peers come up automatically; tier-2 peers are opt-in. geniesim docker up's entrypoint editable-installs every tier-1 peer declared in source/geniesim/pyproject.toml. The asset library geniesim_assets is a tier-1 peer distributed out of band (sourced from HuggingFace / ModelScope), and editable-installed by the entrypoint so every other peer can find it.

Tier-2 peers (geniesim_teleop ยท geniesim_scene_agent ยท geniesim_world) have heavy stacks (VR / LLM / CUDA-ML) and are not auto-installed. Install only the module you need, following its teleop, scene-agent, or World setup guide. World uses a dedicated environment. geniesim bootstrap installs or repairs required peers and skips these optional packages.

geniesim_dream is a separately installed first-class package with its own environments; it is outside the umbrella's Tier-1/Tier-2 extras. See the module tiers.

3.2 ๐Ÿงช Benchmark โ€” debug locally, compete globally on RoboColiseum

Isaac benchmark tasks use geniesim docker / geniesim docker6.2; see the benchmark runtime guide.

flowchart LR
  list(["๐Ÿ“‹ benchmark<br/>list / categories"]) ==> probe(["๐Ÿ”Œ check-<br/>inference"]) ==> run(["๐Ÿงช benchmark<br/>run ยท batch"]) ==> submit(["๐Ÿ† RoboColiseum<br/>leaderboard"])
  classDef step fill:#ede9fe,stroke:#5b21b6,color:#3b0764,font-weight:bold
  class list,probe,run,submit step
  linkStyle 0,1,2 stroke:#5b21b6,stroke-width:2px
Loading

The Genie Sim Benchmark is the engine behind RoboColiseum โ€” and the same tool you use to get ready for it:

  • ๐Ÿ† It's a competition. RoboColiseum ranks embodied-AI policies on four boards (instruction / robust / manip / spatial).
  • ๐Ÿงช Debug locally โ†’ submit remotely. Run the benchmark tasks on your own machine to debug your policy against your own inference server, then launch a model evaluation on the official leaderboard with the same server โ€” no blind submissions.
  • ๐Ÿค– Agent-friendly evaluation submission skills. RoboColiseum provides ready-to-use skills for login, asset and dataset downloads, baseline setup, connecting existing inference services, submitting evaluation jobs and tracking results. AI agents such as Claude Code and Cursor can follow the skills guide to run the workflow.

Run it locally โ€” cat source/geniesim_benchmark/USAGE.md for how to run a benchmark task on your own machine against your inference server (CLI verbs, common commands, and overrides).

3.3 โšก RT Engine โ€” realtime, interactive ROS 2 simulation

flowchart LR
  build(["๐Ÿ”จ ros build<br/>dev"]) ==> scene(["๐ŸŽฌ launch<br/>scene ร— launcher"]) ==> sim(["โšก RT Engine<br/>PhysX ยท Newton ยท mjwarp"]) ==> moveit(["๐Ÿฆพ MoveIt 2<br/>+ WBC RViz"])
  classDef step fill:#dcfce7,stroke:#15803d,color:#14532d,font-weight:bold
  class build,scene,sim,moveit step
  linkStyle 0,1,2 stroke:#15803d,stroke-width:2px
Loading

The RT Engine ships its own SKILLs that walk you through every step โ€” workspace build โ†’ scene launch โ†’ MoveIt + WBC RViz โ†’ bringing your own robot โ†’ teleop โ†’ recording โ†’ debugging. Don't memorize ros2 launch flags โ€” follow the skill:

# inside the container:
geniesim ros build dev
cat source/geniesim_ros/skills/build-workspace/SKILL.md    # `geniesim ros build dev`
cat source/geniesim_ros/skills/launch-scene/SKILL.md       # scene ร— launcher matrix
cat source/geniesim_ros/skills/moveit-wbc/SKILL.md         # MoveIt 2 + WBC RViz for Genie G2
cat source/geniesim_ros/skills/add-robot/SKILL.md          # bring your own URDF / xacro
cat source/geniesim_ros/skills/*/SKILL.md                  # and more

3.4 ๐ŸŒ Generate a 3D world โ€” geniesim_world

flowchart LR
  pano(["๐Ÿ–ผ๏ธ Panorama<br/>(.png)"]) ==> da(["๐Ÿ”ญ DA360<br/>depth"]) ==> sharp(["โœจ SHARP<br/>Gaussians"]) ==> world(["๐ŸŒ 3D world<br/>(PLY ยท Gaussians)"])
  classDef step fill:#ffedd5,stroke:#c2410c,color:#7c2d12,font-weight:bold
  class pano,da,sharp,world step
  linkStyle 0,1,2 stroke:#c2410c,stroke-width:2px
Loading

Turn an equirectangular panorama into an explorable photorealistic 3D world. The generated PLY / Gaussians are consumed by geniesim_world's own tooling. RT Engine scene loading is not supported; see the platform roadmap for proposed integration.

cat source/geniesim_world/skills/generate-world/SKILL.md

3.5 ๐Ÿ’ฌ Generate a scene from a prompt โ€” geniesim_scene_agent

flowchart LR
  nl(["๐Ÿ’ฌ NL prompt<br/>'an apple on the table'"]) ==> assets(["๐Ÿ” search<br/>asset library"]) ==> dsl(["๐Ÿ“ Scene recipe<br/>DSL / direct / hybrid"]) ==> usda(["๐Ÿ—๏ธ scene.usda<br/>+ scene metadata"])
  classDef step fill:#ede9fe,stroke:#6d28d9,color:#3b0764,font-weight:bold
  class nl,assets,dsl,usda step
  linkStyle 0,1,2 stroke:#6d28d9,stroke-width:2px
Loading

Describe a scene in words, search the external asset library, and generate a foreground scene using the bundled Python tools and agent skills. Scene authoring is a standalone workflow; robot task execution uses the independent collection runtime. Optional GPU embeddings support asset retrieval, and rendering uses a configured Isaac Sim environment or the scene-agent image.

python -m pip install -e 'source/geniesim_scene_agent[scene]'
geniesim-scene-agent resources
geniesim-scene-agent assets-root

cat source/geniesim_scene_agent/skills/geniesim-asset-search/SKILL.md
cat source/geniesim_scene_agent/skills/geniesim-scene-generation/SKILL.md
cat source/geniesim_scene_agent/skills/geniesim-scene-to-data/SKILL.md

See the scene-agent guide for scene generation and optional task orchestration, the container guide for rendering, and the example library for task and scene references. The module map lists each package's docs and skills.

3.6 ๐ŸŽฎ Teleoperation โ€” geniesim_teleop

flowchart LR
  up(["๐Ÿณ docker6.2 up"]) ==> collect(["๐ŸŽฎ autoteleop.sh --isaac-version 6.2<br/>VR collection"]) ==> keep(["๐Ÿ’พ keep / discard<br/>(y / n)"]) ==> post(["๐Ÿ“ฆ post-process<br/>โ†’ HDF5 episodes"])
  classDef step fill:#fce7f3,stroke:#be185d,color:#831843,font-weight:bold
  class up,collect,keep,post step
  linkStyle 0,1,2 stroke:#be185d,stroke-width:2px
Loading

Collect demonstration data by teleoperating the robot with a Pico VR headset, then turn the recorded rosbags into HDF5 episodes. Both ordinary Pico and Wuji launchers start Isaac Sim 6.2 in manual collection mode, without automatic evaluation. For simulator startup, build with geniesim docker6.2 build. Once the GUI container is up (geniesim docker6.2 up succeeds), the whole flow is two host-side scripts:

# 1. VR collection โ€” press y (keep) / n (discard) when the collection is done:
./source/geniesim_teleop/scripts/autoteleop.sh --isaac-version 6.2

# 2. Post-process the recording into HDF5 episodes:
./source/geniesim_teleop/scripts/autoteleop_post_process.sh --isaac-version 6.2 <task_name>

Isaac 6.2 mode (--isaac-version 6.2, the default) uses geniesim4 and starts only control processes; provide the simulator separately.

See source/geniesim_teleop/README.md for the controller button map, Pico setup, and task configuration.

3.7 ๐Ÿ“ฅ Data collection โ€” data_collection

flowchart LR
  build(["๐Ÿ”จ data-collection build"]) ==> list(["๐Ÿ“‹ data-collection llm list<br/>pick a task"]) ==> run(["๐Ÿ“ฅ data-collection run<br/>cuRobo auto-planning"]) ==> ep(["๐Ÿ’พ episodes<br/>recording_data/"])
  classDef step fill:#e0f2fe,stroke:#0369a1,color:#0c4a6e,font-weight:bold
  class build,list,run,ep step
  linkStyle 0,1,2 stroke:#0369a1,stroke-width:2px
Loading

Collect automated tasks with cuRobo v2 trajectory planning in an independent Isaac Sim 6.2 image. The host CLI registers the module and locates external assets and task libraries; runtime dependencies are installed in the collection image.

# 1. Register the module and build its complete runtime image:
geniesim data-collection install
geniesim data-collection build    # optional: --china or --mirrors NAME

# 2. Discover LLM tasks, then collect one:
geniesim data-collection llm list --robot g2
geniesim data-collection run <TASK> --scene <SCENE_ID> --run-times 1 --render

# Declarative JSON jobs use a separate entry:
geniesim data-collection run --task-json tasks/my_job.json --headless --standalone

The image defaults to registry.agibot.com/genie-sim/geniesim4-data-collection:latest and extends the dedicated geniesim4:latest base with Isaac Sim 6.2 and ROS 2 Jazzy. Configure external assets with GENIESIM_ASSETS_PATH and task libraries with GENIESIM_LLM_TASK_ROOT or the install command's --assets-root / --task-root options.

Episodes land in source/data_collection/recording_data/. See the collection guide, task configuration, and replay for the runtime and output contracts.

3.8 ๐Ÿง  Agent2Act โ€” LLM-to-action runtime

Agent2Act is a closed-loop framework that connects multimodal LLMs to robot actions in Genie Sim. It supports motion generation, policy backends such as ฯ€0.5 / OpenPI, and automatic routing between execution modes, using visual observations and robot feedback to guide each step.

See source/agent2act/README.md for installation, configuration, and usage instructions.

3.9 ๐ŸŒ™ World-model inference โ€” geniesim_dream

Generate G1/G2 robot videos from recorded actions or a pi0.5 policy loop, with optional VLM evaluation. Dream provides a browser UI and 20 example bundles. It uses separate world-model, policy and evaluator environments, outside the default geniesim bootstrap flow.

See the Dream quick start, installation, and model assets. Its pinned Cosmos Framework and OpenPI dependencies are Git submodules; checkpoints are downloaded separately.

3.10 Post-training server

flowchart LR
  build(["๐Ÿ”จ Build the Docker images"]) ==> shell(["๐Ÿ’ป Enter the container"]) ==> server(["โš™๏ธ Start the server"])
  classDef step fill:#ccfbf1,stroke:#0f766e,color:#134e4a,font-weight:bold
  class build,shell,server step
  linkStyle 0,1 stroke:#0f766e,stroke-width:2px
Loading

Start a GenieSim simulation server for policy training and evaluation. Prepare the host CLI and an editable geniesim_assets installation in the same Python environment (see ยง 3.1). The asset directory is detected and mounted automatically.

Build the GenieSim base from NVIDIA's Isaac Sim 6.2 image, then add the post-training layer:

# On the simulation host, outside containers, from the repository root:
geniesim docker6.2 build
geniesim docker --profile post-train build
geniesim docker --profile post-train up --headless
geniesim docker --profile post-train into
# Inside the post_train_server container, from /workspace:
geniesim post-train server start --count 1 --base-port 5601 --devices 0

This starts one server on GPU 0, listening on TCP port 5601. A client selects the task and initializes the simulation.

For training and evaluation with RLinf, see the RLinf GenieSim guide.

4. ๐Ÿ“œ Roadmap & Updates

The arc of the platform โ€” what shipped when, and what's coming next.

flowchart LR
  v2["<b>v2.x ยท 2025</b><br/>Challenge tasks<br/>2ร— sim speed"]
  v30["<b>v3.0 ยท Jan 2026</b><br/>Isaac Sim 5.1<br/>Genie G2 ยท 3DGS"]
  v31["<b>v3.1 ยท Apr 2026</b><br/>Genie Sim World<br/>RLinf"]
  v32["<b>v3.2 ยท Jun 2026</b><br/>๐Ÿงž CLI ยท ๐Ÿค– SKILLs<br/>โšก RT Engine"]
  v35["<b>v3.5.0 ยท Nov 2, 2026</b><br/>Benchmark-focused release<br/>Isaac Sim 5.1 + 6.2 combo"]
  v36["<b>v3.6.0</b><br/>Isaac Sim 6.2 runtime<br/>Reproducible dependency locks"]
  next["<b>v4.0 line</b><br/>Dedicated Isaac Sim 6.2<br/>RT Engine qualification"]
  v2 --> v30 --> v31 --> v32 --> v35 --> v36 --> next

  classDef done fill:#d1f0d8,stroke:#1f6f3d,color:#0b3b1f
  classDef current fill:#cfe2ff,stroke:#084298,color:#084298,stroke-width:2px
  classDef future fill:#fff3cd,stroke:#856404,color:#5c4400
  class v2,v30,v31,v32,v35 done
  class v36 current
  class next future
Loading

๐Ÿ“œ Release log

v3.6.0 โ€” Isaac Sim 6.2
  • ๐Ÿณ Dedicated Isaac Sim 6.2 runtime โ€” system Python uses the bundled SDK in geniesim4:latest; benchmark, IK/FK and teleoperation workflows target this runtime.
  • ๐Ÿ”’ Reproducible dependencies โ€” pinned uv, compact locks with artifact hashes, and SDK/provider validation for the dedicated 6.2 image.
  • ๐Ÿ”– Package version โ€” all eight first-class Genie Sim distributions and 15 owned ROS manifests share VERSION=3.6.0. Image-family names, dependency versions and independent modules retain their own versions.
[11/2/2026] v3.5.0 โ€” Benchmark release (5.1 + 6.2 combo now E.O.L.)
  • ๐Ÿงช Extended benchmark suites โ€” fine manipulation, dexterous manipulation, whole-body coordination, and cross-embodiment tasks, plus 25 unscored open-world sandbox scenarios.
  • โš™๏ธ Task-selected runtimes โ€” direct Isaac evaluation, managed Newton wok evaluation, and native MuJoCo Rubik execution with separate observation rendering.
  • ๐Ÿณ Benchmark combo runtime โ€” geniesim3:latest used bundled Isaac Sim 5.1 and the system-Python Isaac Sim 6.2 co-install; that combo is now deprecated (E.O.L.). The dedicated geniesim4:latest image remains on the separate 4.0 line.
  • ๐ŸŒ™ Genie Sim Dream โ€” G1/G2 world-model inference, policy integration, visual evaluation, and a browser UI in dedicated environments.
  • ๐Ÿ“ฅ Collection and post-training workflows โ€” independent Isaac Sim 5.1/cuRobo v2 collection and replay, scene-agent authoring workflows, and RLinf server/client integration.
  • ๐Ÿ”– Package version at release โ€” all eight first-class Genie Sim distributions and owned ROS package metadata shared VERSION=3.5.0. This package version did not relabel the dedicated 4.0 image line. Dependencies, recording formats, and independently maintained modules kept their own versions.

See the module index and benchmark task commands for workflow details.

[6/25/2026] v3.2.0 โ€” Genie Sim RT Engine + Agentic CLI/SKILLs
  • ๐Ÿ† RoboColiseum โ€” opened RoboColiseum: submit a policy and get it scored on the official leaderboard. Added a new spatial board (alongside instruction / robust / manip), and the robust board now aggregates and scores tasks per perturbation type (instruction, robot pose, background, image quality, camera position). The full pipeline โ€” download datasets, fetch baseline weights, run an example leaderboard service, submit & track โ€” ships as agent-friendly one-click SKILLs under source/geniesim_benchmark/skills/robocoliseum/.
  • ๐Ÿงž geniesim CLI โ€” one command for docker, ROS 2 builds, bootstrap, status, doctor, deploy; bash/zsh completion; standalone PEP 517 / PEP 621 wheel.
  • ๐Ÿค– Agent SKILLs โ€” self-contained SKILL.md recipes (run-benchmark, check-inference, geniesim-scene-generation, geniesim-asset-search, geniesim-scene-to-data, run-teleop) that Claude Code and other agents invoke to drive the simulator end-to-end.
  • โšก Genie Sim RT Engine (realtime, interactive) โ€” geniesim_ros ships physics + render + robot as a first-class ROS 2 node sharing one sim_time. Multiple physics backends (Isaac Sim PhysX, Isaac Sim Newton, Kit-free Newton-standalone) with cloth / soft-body on the Newton path.
  • ๐Ÿฆพ Customizable robots โ€” Genie G2 family (Tier 1, continuously maintained) with arm ร— gripper matrix, drop-in MoveIt 2 + WBC RViz, three IK plugins; reference URDF for Franka, UR5, Aloha, ARX, Agilex; bring-your-own via xacro + offline mesh-prep tools.
  • ๐ŸŽฌ New demos โ€” pick-and-place (scene_pnp_g2_op) and whole-body control (scene_wbc_g2_sp).
  • ๐Ÿ› ๏ธ Engine internals โ€” AS3 asset layout (per-robot robot.usda + payloads/Physics/{physics,physx,mujoco}.usda); IsaacSim 5.1 + 6.0 share one CLI surface; fisheye camera + de-skewed rotary lidar.
  • ๐Ÿ”— Module dependency DAG โ€” auto-generated from each peer's pyproject.toml, rendered in source/README.md; enforced by geniesim tool deps-dag in CI.
[4/8/2026] v3.1
  • ๐ŸŒ Release Genie Sim World โ€” a multimodal spatial world model for 3D world generation.
  • ๐Ÿงช Update new benchmarks for instruction following, spatial understanding, manipulation skills, robustness, and sim2real.
  • ๐ŸŽ“ Support human-in-the-loop and distributed reinforcement learning pipeline of RLinf.
[1/7/2026] v3.0
  • ๐Ÿ†™ Update Isaac Sim to v5.1.0 and support RTX 50series graphic card.
  • ๐Ÿฆพ Provide USD and URDF files of Agibot Genie G2 robot and support whole body control.
  • ๐Ÿ›ฐ๏ธ Support 3DGS-based scene reconstruction and convert output to USD format for application in Isaac Sim.
  • ๐Ÿ“ฆ Release synthetic dataset and corresponding data collection pipeline.
  • ๐Ÿ’ฌ Add LLM-based features to generate scenarios, task instructions and evaluation configurations.
[7/14/2025] v2.2
  • ๐Ÿ“Š Provide detailed evaluation metrics for all Agibot World Challenge tasks.
  • โš™๏ธ Add automatic evaluation script to run each task multiple times and record score of all steps.
[6/25/2025] v2.1
  • โž• Add 10 more manipulation tasks for Agibot World Challenge 2025 including all simulation assets.
  • ๐Ÿ“ฆ Open-source synthetic datasets for 10 manipulation tasks on Huggingface.
  • ๐Ÿ”Œ Integrate UniVLA policy and support model inference simulation evaluation.
  • ๐Ÿงฎ Update IK solver sdk which supports cross-embodiment IK solving for other robots.
  • โšก Optimize communication framework and improve simulation running speed by 2ร—.
  • ๐Ÿงช Update automatic evaluation framework for more complicated long-range tasks.

5. ๐Ÿ“š Documentation

5.1 ๐Ÿ“– Documentations

Please refer to these links to download assets and dataset:

5.2 ๐Ÿ’ฌ Support

5.3 ๐Ÿ“„ License and Citation

All the data and code within source/geniesim_* and source/data_collection are under Mozilla Public License 2.0 unless licensed specifically. The source/scene_reconstruction project contains code under multiple licenses, for complete and updated licensing details, please see the LICENSE files

Please consider citing our work either way below if it helps your research.

@misc{yin2026geniesim30,
  title={Genie Sim 3.0 : A High-Fidelity Comprehensive Simulation Platform for Humanoid Robot},
  author={Chenghao Yin and Da Huang and Di Yang and Jichao Wang and Nanshu Zhao and Chen Xu and Wenjun Sun and Linjie Hou and Zhijun Li and Junhui Wu and Zhaobo Liu and Zhen Xiao and Sheng Zhang and Lei Bao and Rui Feng and Zhenquan Pang and Jiayu Li and Qian Wang and Maoqing Yao},
  year={2026},
  eprint={2601.02078},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2601.02078},
}

5.4 ๐Ÿ”— References

  1. PDDL Parser (2020). Version 1.1. [Source code]. https://github.com/pucrs-automated-planning/pddl-parser.
  2. BDDL. Version 1.x.x [Source code]. https://github.com/StanfordVL/bddl
  3. CUROBO [Source code]. https://github.com/NVlabs/curobo
  4. Isaac Lab [Source code]. https://github.com/isaac-sim/IsaacLab
  5. Omni Gibson [Source code]. https://github.com/StanfordVL/OmniGibson
  6. The Scene Language [Source code]. https://github.com/zzyunzhi/scene-language
  7. COAL [Source code]. https://github.com/coal-library/coal
  8. OCTOMAP [Source code]. https://github.com/OctoMap/octomap
  9. PINOCCHIO [Source code]. https://github.com/stack-of-tasks/pinocchio
  10. URDFDOM [Source code]. https://github.com/ros/urdfdom
  11. LIBCCD [Source code]. https://github.com/danfis/libccd
  12. LIBMINIZIP [Source code]. https://github.com/switch-st/libminizip
  13. LIBODE [Source code]. https://github.com/markmbaum/libode
  14. LIBURING [Source code]. https://github.com/axboe/liburing
  15. MuJoCo [Source code]. https://github.com/google-deepmind/mujoco
  16. Wuji SDK [Source code]. https://github.com/wuji-technology/wuji-sdk โ€” MIT license

About

Simulation Platform from AgiBot

Resources

Stars

1.7k stars

Watchers

11 watching

Forks

Releases

Packages

Contributors

Languages