NVIDIA Isaac Integration: What's New in v2.0

Today we're releasing Nepher v2.0, featuring deeper integration with NVIDIA Isaac Sim and Isaac Lab. This release represents months of engineering work to make simulation-first robotics development faster, more capable, and more accessible than ever.

Key Highlights

• Native Isaac Sim 4.5 support with all the latest physics improvements
• 3x faster training throughput on supported GPU configurations
• New asset import pipeline supporting OnShape, Fusion 360, and SolidWorks
• Improved domain randomization API
• Real-time policy visualization and debugging tools

Performance Improvements

Performance has been our top priority for v2.0. Through aggressive optimization of the simulation-to-training pipeline, we've achieved up to 3x faster training throughput on RTX 4090 hardware and 2.5x on H100 systems.

These improvements come from several sources: optimized memory management, reduced CPU-GPU synchronization overhead, and improved batching strategies. The result is that training runs that previously took 24 hours now complete in 8 hours or less.

New Asset Import Pipeline

One of the most-requested features is here: native support for popular CAD formats. The new asset import pipeline accepts OnShape, Fusion 360, and SolidWorks files, automatically converting them to USD with proper joint configuration and physics properties.

The pipeline handles common CAD-to-simulation challenges automatically:

• Joint identification and configuration from CAD assemblies
• Mass property calculation from solid geometry
• Collision mesh simplification for performance
• Material property assignment from CAD metadata

Enhanced Domain Randomization

Domain randomization has been completely revamped with a new declarative API. Instead of writing complex randomization code, you can now specify randomization parameters through a simple YAML configuration or our visual configuration tool.

The new API supports:

• Parameter randomization with custom distributions
• Correlated parameter randomization
• Curriculum-based randomization scheduling
• Environment-specific randomization profiles

Visualization and Debugging

Understanding what your policy is doing has never been easier. The new visualization tools provide real-time insight into:

• Policy attention maps for vision-based policies
• Reward component breakdowns
• Action distributions and their evolution
• State trajectories with anomaly detection
• Value function gradients and uncertainty estimates

Improved Sim-to-Real Pipeline

We've also improved the sim-to-real pipeline with new tools for hardware validation. The platform now supports direct streaming of trained policies to physical robots through standard ROS 2 interfaces, with built-in safety monitors and emergency stop functionality.

Breaking Changes

v2.0 includes some breaking changes from v1.x. The most significant changes are:

• New configuration file format (migration tool provided)
• Updated Python API for environment registration
• Revised tensor layout for observation and action spaces

We've published a comprehensive migration guide in our documentation. For most users, migration takes less than an hour.

Getting Started with v2.0

v2.0 is available now to all Nepher users. Existing users can upgrade through the platform settings, and new users will receive v2.0 by default. We recommend reviewing the migration guide and trying the new features in a development environment before updating production workflows.