AgibotTech/genie_sim

What it solves

Genie Sim is a comprehensive simulation platform designed to accelerate the development of embodied intelligence (robotics). It reduces the reliance on physical hardware by providing tools for high-fidelity environment reconstruction, synthetic data collection, and automated evaluation of robot policies.

How it works

The platform integrates 3D reconstruction and visual generation to create realistic simulation environments. It uses LLM-driven technology to generate diverse scenes and evaluation configurations from natural language prompts. The system is built around a ROS 2 native engine that supports multiple physics backends (such as Isaac Sim PhysX and Newton) and includes a CLI for managing the Docker lifecycle and workspace builds.

Who it’s for

It is intended for developers and researchers working on embodied AI, specifically those building and testing robot manipulation and loco-manipulation policies.

Highlights

• LLM-Driven Scene Generation: Creates simulation scenes instantly via conversational interaction and natural language prompts.
• Genie Sim Benchmark: A standardized tool with over 100,000 scenarios and 200+ tasks to establish capability profiles for AI models.
• High-Fidelity Assets: Includes over 5,000 validated 3D assets across retail, industrial, catering, home, and office environments.
• Synthetic Data Pipeline: Provides over 10,000 hours of synthetic data and tools for both VR teleoperation and automated data collection using cuRobo.
• Agent-Ready SKILLs: Includes machine-readable recipes (SKILL.md) that allow AI coding agents (like Claude Code or Cursor) to operate the simulator without human intervention.
• Sim-to-Real Transfer: Enables zero-shot transfer of models trained on synthetic data to real-world robots.