Genesis-Embodied-AI/genesis-world
Genesis World
Genesis World is an embodied AI simulation platform designed for training robotic agents through physics-based interactions. It centers on a multi-physics simulation engine that integrates rigid body, particle, and finite element method dynamics, supported by a parallel simulation kernel compiler that translates Python functions into optimized GPU and CPU kernels.
The platform features a photorealistic robot renderer that utilizes path-tracing and Gaussian Splatting to generate synthetic training data. It includes a domain randomization framework to vary lighting and physical parameters across parallel instances and a robotics sensor simulator for virtual depth cameras, lidar, IMUs, and tactile sensors.
The system provides broader capabilities for robotic agent control via kinematics solvers and joint controllers, as well as tools for simulation asset parsing and the attachment of virtual sensors to robotic links using dynamic transforms.
Features
- Embodied AI Platforms - Provides an integrated platform for training autonomous agents to learn physical interactions and navigation.
- Multi-Physics Simulation Engines - Provides a multi-physics simulation engine integrating rigid body, FEM, and particle dynamics for complex robotic interactions.
- Photorealistic Scene Synthesis - Generate path-traced images of objects and terrain using GPU assets to feed training loops for high-fidelity visual data.
- Physics and Simulation - Integrates rigid body, FEM, and particle dynamics into a shared scene for complex physical interactions.
- Multi-Physics Integration - Combines rigid body, finite element method, and particle dynamics within a single unified simulation environment.
- Robotics Simulators - Provides a comprehensive robotics simulator for testing autonomous agents with simulated depth cameras, lidar, and tactile sensors.
- Parallelized Robotics Simulations - Uses parallel computing kernels to run large-scale robotic training loops and physics calculations.
- Domain Randomizations - Varies lighting and environmental parameters across parallel simulations to improve AI model generalization.
- Robotic Sensor Simulation - Creates virtual depth cameras, lidar, and tactile sensors to provide real-time environmental feedback for robotic controllers.
- Robotics Perception Simulators - Simulates depth cameras, IMUs, lidar, and tactile sensors to provide real-time environmental feedback to autonomous agents.
- Physics Simulation - Models complex interactions between rigid bodies, particles, and soft bodies to test robotic hardware.
- Synthetic Data Renderers - Ships a GPU-accelerated renderer using path-tracing and Gaussian Splatting to generate photorealistic synthetic training data.
- Robotics and Control - Manages autonomous agent behavior using kinematics solvers, joint controllers, and specialized hardware interfaces.
- GPU-Accelerated Compilers - Translates high-level Python functions into optimized machine code for parallel execution on GPUs.
- Compiler Backends - Lowers high-level simulation code to target various hardware backends, including GPU and CPU architectures.
- JIT Kernel Compilers - Implements a JIT kernel compiler that translates Python functions into optimized GPU and CPU kernels for parallel simulation.
- Python GPU Kernels - Translates Python functions into optimized hardware kernels for high-performance parallel computing across graphics backends.
- Parallel Scene Simulations - Implements parallel scene simulations to accelerate data collection across multiple hardware threads.
- Parallel Rendering - Generates multiple simultaneous environment instances with varied lighting to improve agent generalization.
- Hardware-Threaded Loop Distribution - Accelerates data-parallel computations by distributing top-level simulation loop iterations across hardware threads.
- High-Quality Scene Rendering - Generates high-quality photorealistic visuals and rasterized views to simulate robotic camera sensor output.
- Gaussian Splatting - Utilizes Gaussian Splatting to create photorealistic 3D backgrounds from real-world light field data.
- Path Tracing Implementations - Produces photorealistic frames for training loops using GPU-accelerated path-tracing and light transport simulation.
- Rendering Randomizers - Renders multiple parallel environments with unique lighting conditions to increase the diversity of training data.
