The field of embodied intelligence is witnessing a fierce debate over the best approach to training robots for industrial applications. One faction advocates for simulation-based training, leveraging structured environments to generate synthetic data, while the opposing view emphasizes the necessity of real-world data to handle complex physical interactions and unpredictable scenarios. Key players include NVIDIA, DeepMind, and Intrinsic, each with unique strategies and technologies.
NVIDIA's Omniverse platform and Isaac Sim engine exemplify the simulation approach, enabling comprehensive digital twins of factories for training and optimization. Their collaboration with BMW on a digital twin project in Hungary showcases the potential of synthetic data in logistics and robotic movements. However, challenges remain in achieving the necessary fidelity for force control and physical interactions, prompting NVIDIA to seek partnerships with companies like Hexagon Robotics.
Conversely, DeepMind's use of the MuJoCo physics engine has demonstrated that pure simulation can achieve industrial-grade precision in specific tasks, such as sorting with known rigid models. Yet, this method's effectiveness is limited to scenarios with minimal contact and force control. Intrinsic aims to transform simulation into a comprehensive development tool for industrial robots, focusing on lowering barriers for small manufacturers. The ongoing challenge of the SIM2REAL gap remains a critical factor in the success of these approaches.
Editor's Note
The ongoing debate between simulation and real-world training methods highlights a significant shift in robotics development strategies. Companies are increasingly focusing on creating robust infrastructures that can support diverse applications across various industrial sectors.
Leave a comment