# Genie Sim: Open Source Simulation for Robot Evaluation > AgibotTech/genie_sim is a Python and ROS 2 simulation platform for reconstructing environments, generating scenes, collecting synthetic robot data and evaluating embodied intelligence models. Its main adoption cost is a container-based Isaac Sim and ROS 2 setup rather than a lightweight pip-only toolchain. **AgibotTech/genie_sim** — Simulation Platform from AgiBot - Repository: https://github.com/AgibotTech/genie_sim - Stars: 1,504 · Forks: 135 - Language: Python - License: NOASSERTION - Published: 2026-10-16 · Updated: 2026-10-16 · Language: en - Canonical page: https://hysenlabs.com/projects/agibottech-genie-sim ## Genie Sim's Role in Robot Development Genie Sim addresses a specific problem in embodied intelligence: developers need repeatable simulated environments in which robot policies can be exercised across many tasks, scenes, sensors and evaluation conditions. The platform is aimed at developers working with manipulation, locomotion-manipulation and robot operation data. According to the README, it combines environment reconstruction, scene generalization, data collection and automated evaluation rather than acting as only a physics simulator or only a dataset viewer. The scope is broad. Genie Sim includes a benchmark with more than 200 tasks and more than 100,000 scenarios, an LLM-driven scene and evaluation configuration workflow, synthetic data tooling, teleoperation and ROS 2 integration. It also publishes links to asset and dataset repositories on ModelScope, as well as a separate Genie Sim World component for generating photorealistic 3D worlds from diverse inputs. Those numbers and capabilities are project claims in the README, not independently verified measurements. This distinction matters for adoption. A team looking for a small simulator with a few hand-written scenes may find the platform too broad. A team building robot learning pipelines can treat the project as a container workspace that joins assets, ROS 2 packages, interactive control and evaluation. The repository's own organization supports that interpretation: the top level contains `source/`, `docker/`, `docs/`, `3rdparty/` and agent instructions, rather than presenting a single library with one import path. ## ROS 2, Containers and the Genie Sim CLI The documented first-use path is a host-side CLI installation followed by a Docker image build. The CLI is packaged as a standalone PEP 517 and PEP 621 wheel, and the README identifies the package directory as `source/geniesim_cli/`. The build command is described as creating an image containing Isaac Sim 5.1 and Jazzy. Once the container starts, its entrypoint performs editable installs for tier-1 packages in the bind-mounted workspace. The CLI provides the lifecycle commands used in the quick-start sequence. `geniesim status` performs a per-distribution health probe. `geniesim doctor` is documented as diagnosing and repairing conditions such as `rosdep` and environment configuration. `geniesim bootstrap` reinstalls tier-1 peers and supports opt-in tier-2 packages. The README also describes Docker lifecycle management, ROS 2 workspace builds, asset and scene bootstrap, and pure-Python wheel deployment as part of the CLI's scope. The interactive path uses `geniesim docker into`, which opens a shell inside the container. The README's flow places host installation before image build, image startup before peer installation, and shell access after startup. That sequence is useful because it keeps the initial setup visible at the host boundary while allowing package-level work inside the simulation workspace. ## Installing Genie Sim and Checking the Stack The documented host-side setup is: ```bash pip install -e source/geniesim_cli/ geniesim docker build geniesim docker up geniesim docker into ``` After entering the container, check the installed distributions and run the documented setup and diagnosis commands. The README says `status` reports per-distribution health, `doctor` diagnoses and repairs items such as `rosdep` and environment settings, and `bootstrap` reinstalls tier-1 peers with opt-in tier-2 support. ## Genie Sim Real-Time Simulation Through ROS 2 Genie Sim RT Engine is the interactive and real-time part of the platform. The README states that `geniesim_ros` brings physics, rendering and the robot into ROS 2, allowing teleoperation, MoveIt and `ros2_control` to operate on one shared `sim_time`. This is a concrete architecture claim: the simulation and ROS-facing components are intended to share simulation timing rather than run as unrelated processes with different clocks. Physics backends can be selected per scene. The documented choices are Isaac Sim PhysX and Newton-standalone. The Newton path is also associated with cloth and soft-body support. This distinction may affect scene selection because the README does not state that every feature is available on every backend. The repository also packages `SKILL.md` recipes for benchmarking, inference checks, scene generation, asset search and teleoperation. The README says these recipes can be invoked by Claude Code and other agentic coding agents, while `AGENTS.md` files provide the boot sequence, module map, architecture notes and commands. For a human user, the documented reading path starts with the root `AGENTS.md`, then `source/AGENTS.md`, a package-level `AGENTS.md`, and finally the relevant `skills//SKILL.md` recipe. This is useful when the CLI is used for a specific benchmark or scene because the repository provides machine-readable signposts instead of requiring every workflow to be discovered from the top-level README. ## Genie Sim Assets, Scene Generation and Synthetic Data The platform separates reusable assets from generated scenes and data. The README advertises 5,140 validated 3D assets across retail, industry, catering, home and office. It also describes a 3DGS-based reconstruction pipeline that combines reconstruction with a visual generative model to produce simulated environments with high-precision meshes. The associated data is linked through the project's ModelScope pages. Scene creation is described as LLM-driven. Natural language is used to generate and generalize simulation scenes through conversational interaction, and the benchmark workflow uses an LLM to produce task instructions and evaluation configurations. The README also describes more than 10,000 hours of open synthetic data across 200-plus locomotion-manipulation tasks, with multi-sensor streams and multiple dimensions of variation. The data-generation toolkit supports low-latency teleoperation and automated data programming, and it includes an error-recovery mechanism. Those features point toward repeatable collection rather than isolated demonstrations. However, the README does not document a general rollback procedure for generated scenes or collected datasets. It also does not specify the hardware profile needed to process the full asset set or synthetic-data workload. A prospective user must therefore distinguish between a documented CLI operation and an undocumented capacity requirement. ## Genie Sim Benchmark Claims, Limits and License Boundary The benchmark is positioned as a capability profile rather than a single task runner. The README says evaluation criteria cover manipulation skills, cognitive comprehension and task complexity, while its stated scale is 100,000-plus scenarios. It also reports a discrepancy below 10% between simulation and real-world test results, but the README excerpt does not identify the datasets, model families, statistical method or conditions behind that figure. The claim should therefore be treated as a project-reported result, not as a general guarantee for every policy or task. Genie Sim is not a replacement for real-robot validation. Simulation-to-real behavior can differ, and the README does not document hardware calibration procedures, simulator settings, sensor noise policies or a complete process for reproducing the stated discrepancy. A team evaluating a safety-critical controller should not infer deployment readiness from the synthetic-data claim alone. Maintenance information is limited to repository facts. The repository is not archived, and the last push was on 2026-09-07. The release list includes `v3.2.0` on 2026-08-04. That does not establish that every ROS 2 package or external asset is updated together. The license is listed as Other because GitHub cannot classify the custom `LICENSE` file. Users must read that file before redistribution, modification, internal deployment or combination with other software. No legal conclusion can be drawn from the repository description alone. ## Conclusion Genie Sim fits robotics teams that need a reproducible simulation workspace for manipulation research, synthetic data collection or model evaluation, and that can operate Isaac Sim 5.1 and ROS 2. It is a poor fit for a quick CPU-only experiment or for teams that require a clearly classified standard software license. Before adoption, inspect the custom LICENSE file, confirm the hardware and container requirements, and run `geniesim docker up`, followed by `geniesim status` and `geniesim doctor`, in the intended environment. ## FAQ ### What is Genie Sim used for? Genie Sim is an AgiBot simulation platform for environment reconstruction, scene generalization, robot data collection and automated evaluation of embodied intelligence models. It combines 3D assets, ROS 2 simulation, teleoperation and benchmark tooling. ### How do I install Genie Sim? The README documents installing `source/geniesim_cli/` on the host, building the Isaac Sim 5.1 plus Jazzy image with `geniesim docker build`, starting it with `geniesim docker up`, and entering it with `geniesim docker into`. It then shows `geniesim status`, `geniesim doctor` and `geniesim bootstrap` as the initial stack checks and setup commands. ### Does Genie Sim support real-time ROS 2 robot control? Yes. The README says `geniesim_ros` connects physics, rendering and the robot to ROS 2, with teleoperation, MoveIt and `ros2_control` sharing `sim_time`. It documents Isaac Sim PhysX and Newton-standalone as selectable physics backends, with cloth and soft-body support described for the Newton path. ## Sources - [AgibotTech/genie_sim on GitHub](https://github.com/AgibotTech/genie_sim) - [Issues](https://github.com/AgibotTech/genie_sim/issues) - [README](https://github.com/AgibotTech/genie_sim/blob/main/README.md) - [Releases](https://github.com/AgibotTech/genie_sim/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/agibottech-genie-sim