#USA #San Jose #Nvidia GTC #XGRIDS #Real2Sim

XGRIDS at NVIDIA GTC 2026: Bridging Real and Virtual Worlds with Real2Sim

At the recent NVIDIA GTC 2026, held from March 16 to 19 in San Jose, XGRIDS unveiled its groundbreaking project, Real2Sim. This initiative focuses on enabling robots to function reliably in real-world scenarios by training them in environments that accurately reflect physical realities. The central question addressed by Sunny Liao, the director of XGRIDS, during a startup pitch, was how to design training environments that genuinely mirror real-world conditions. The solution proposed involves a pipeline based on real data, integrating LiDAR and computer vision to achieve multimodal spatial awareness combined with detailed 3D reconstructions.

Advantages of Real2Sim Approaches

By taking this innovative approach, as opposed to manual 3D modeling, organizations can benefit from:

• - Cost Reduction: Building high-precision environments becomes more affordable, which is crucial for businesses operating on tight budgets.
• - Continuous Updates: As physical spaces evolve, the training models can be continuously updated to reflect these changes, ensuring the robots are training on current data.
• - Real-World Functionality: The simulations remain closer to actual environments, which enhances their reliability when applied in real-world applications.

Developers at the GTC emphasized that this methodology opens more practical paths for training and validating robots, allowing them to adapt dynamically to shifting conditions.

Spatial Intelligence for Physical AI

Beyond its startup pitch, XGRIDS's solutions were highlighted at various presentation areas throughout the GTC. Notably, during the NVIDIA Robotics session, the company demonstrated how their solution could be applied to embodied AI systems. With the integration of spatial perception and modeling on quadrupedal robots, these machines can continually map and comprehend their surroundings. They utilize comprehensive 3D spatial structures for path planning, behavioral decision-making, and task execution, moving beyond reliance on local sensors merely for immediate obstacle avoidance.

This demonstration showcases how spatial intelligence can be effectively integrated into embodied AI systems. Robots can harness complete 3D environments for their operational needs, which could significantly enhance their effectiveness in varied applications.

Full Real2Sim Workflow Overview

During a presentation with Amazon Web Services (AWS), XGRIDS outlined a complete Real2Sim workflow—from data collection to world model generation and training within simulated environments. This exposure allowed attendees to grasp the full potential of XGRIDS's innovative solutions.

Looking Ahead

XGRIDS's long-term goal remains steadfast—constructing an infrastructure for spatial intelligence that translates real environments into world models that AI systems can comprehend, utilize for insights, and apply for training. The GTC 2026 marked another significant step towards integrating this work into the physical AI ecosystem.

As embodied AI systems increasingly transition from laboratory settings to warehouses, urban environments, and construction sites, the demand for precise, scalable digital representations of surroundings will continue to grow. XGRIDS is at the forefront, developing the capture-to-simulation layer that makes this transformation possible, ensuring robots can thrive in real-world settings.