Genesis

Genesis v1.2.3: More Realistic Friction and Rendering

Genesis is a robot simulation platform. This update makes the physics world more realistic: robots no longer fake-slip when pushing boxes or screwing, making grasping and dragging more realistic. The renderer now handles rigid, soft, and articulated bodies in one pass, eliminating separate pipelines. LiDAR sensors have a new 'distances-only' mode that skips complex material interactions, significantly reducing CPU load. It also fixes bugs causing model twisting and phantom collisions, making lar • Introduced an elliptic friction cone model for more accurate static friction, preventing robots from fake-slipping when pushing boxes or screwing, making grasping and dragging more realistic. • The batch renderer now supports rendering rigid, soft, and articulated bodies together in a single pass, eliminating the need for separate rendering pipelines. • Raycasting sensors like LiDAR and depth cameras added a 'distances-only' mode, skipping complex material interactions to significantly cut CPU load for large-scale parallel simulations. • Added an option to ignore the default ground plane, making it easier to create custom terrains like slopes and pits, or run zero-gravity floating experiments.

Genesis Simulator v1.2.2: Tactile Sensors for Dexterous Hands and Collision Detection Overhaul

Genesis Simulator v1.2.2 makes robot training more realistic. Dexterous hands now have tactile sensors that simulate real-world imperfections like dead zones, making RL training more realistic. It also completely solves thin-shell tunneling and ghost collisions for complex shapes. If you are running large-scale reinforcement learning training, computation speed and stability have seen significant improvements. • Added tactile sensors to dexterous hands that simulate real-world imperfections like dead zones, making RL training more realistic. • Upgraded collision detection for complex shapes, completely fixing thin-shell tunneling and spurious contacts. • Optimized the no-slip calculation for grasping; it adds less than 20% overhead but makes holding objects much more stable. • Fixed data access bugs for ultra-large batch simulations and accelerated contact-island constraint solving.

Real2Sim2Real for Vision-Language-Action Manipulation: An AMD ROCm-Based Pipeline

A pipeline for vision-language-action (VLA) manipulation using Real2Sim2Real transfer, built on AMD ROCm for efficient simulation-to-real deployment.