I attended NVIDIA GTC 2025 for a chance to learn about new developments in embodied AI, robotics and large scale synthetic data.
One of the most interesting talks I attended was "The Embodied End-to-End VLA Model Driven by Synthetic Big Data" by He Wang, CTO at Galbot. This talk focused on how vision-language-action (VLA) models can make use of synthetic data to solve issues with sparse data in robotics.
Interesting points covered:
ā£ Using synthetic data for embodied AI ā Learn how large-scale synthetic data can be used to pretrain embodied intelligence models without relying on costly data.
ā£ Sim2Real ā Learn how to train models using simulation to apply robotic capabilities in real world environments.
ā£ VLA Models ā How to combine vision, language and action to create intelligent robotic models.
Examples of robotic models ā Examples included: GraspVLA and Uni-NaVid.
This helped broaden my knowledge of how synthetic data, simulation environments and multimodal learning are being used to develop intelligent robotic models.
It also increased my curiosity around using synthetic data to drive AI models, model generalisation and the challenges faced when developing intelligent models.
Overall it was a great learning opportunity at the intersection of AI, Robotics and next generation computing.