Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.
Engineering & simulationNvidia Isaac Manipulator
Intrinsic and Nvidia combined foundation models with GPU-accelerated libraries to enable robots to perform multiple tasks and handle exceptions in industrial automation.
“Romance are only as smart as what they're able to do. And so with the Isaac manipulator that invented released in March. Uh, it combines these, uh, national models and modular, GPU accelerated libraries, to really help these industrial automation companies, filled, scalable, repeatable workflows.”
Engineering & simulationNvidia Project Groot
Nvidia is developing a large multimodal foundation model to provide a standard baseline for robotic decision-making and motion, allowing developers to build on a common platform rather than creating unique models for every application.
“The concept behind root. I think has a similar idea in the AI space, the foundational model, the concept of what they're calling it, a large multimodal model, the ability for systems to use lots of different inputs and outputs and to take an input and deliver the output signals that can drive the robot to make decisions to make motion to whatever it's entitled to do.”
Engineering & simulationReinforcement learning models
Agility Robotics uses reinforcement learning in simulation to explore the unique dynamics of their robot for low-level control, as there is no source of real-world data for these specific compliance and force behaviors.
“So, the way we see it at the low level for the control of the robot, you have to do that as a simulation. Um, because there's no source of data for it. And you need to kind of explore the unique dynamics of the robot. And so that's a great place for like reinforcing learning models and things like that.”
Engineering & simulationDiffusion-based policy
Agility Robotics is developing a diffusion-based policy to act as a planner in the middle ground between low-level control and high-level language model instructions, requiring significant training data from learned behaviors.
“And then there's this middle ground where, you know, we probably need a ton of training data, maybe from our learned bottles, maybe from our model-based, you know, control models, uh, uh, uh, and all that sort of thing to then create some sort of uh, diffusion-based policy that's kind of like a planner.”