Nvidia and Apptronik bundle artificial intelligence with humanoid robot

Nvidia and Apptronik enter partnership. Apptronik is integrating an AI model from Nvidia for training robots with its humanoid robot Apollo. Mercedes-Benz is currently testing the robot in a factory, where Apollo brings parts to the production line and inspects components.

Apptronik is integrating a general-purpose foundation model for training robots that is part of Project GR00T. This project was announced by Nvidia earlier this month and includes multiple foundation models. GR00T stands for Generalist Robot 00 Technology. It consists of what are called foundation models, deep learning algorithms trained on a very broad dataset. This allows the algorithms to be applied in a wide range of scenarios. Foundation models play an important role in, among others, OpenAI's well-known ChatGPT.

Nvidia aims to use Project GR00T to accelerate the achievement of breakthroughs in robot development and embodied AI. Embodied AI is a term that describes robots that are controlled using artificial intelligence (AI). Among other things, they can interact with and learn from their environment.

Robots using Project GR00T can, among other things, understand natural language and mimic movements by observing human actions. In this way, among other things, they can quickly gain the coordination and agility to navigate, interact with and adapt to the physical world.

Project GR00T was introduced at Nvidia's recent GTC 2024 AI conference. Here, Jensen Huang, Nvidia's top executive, showed several robots using Project GR00T. One is Apollo from Apptronik, a humanoid robot about 172 cm tall and weighing about 72 kg. The robot can lift a weight of almost 25 kg.

Apollo is equipped with a system that adapts the forces exerted by the robot to its environment. This system ensures that the robot can work safely with humans, in a manner very similar to collaborative robots. The safety zone Apollo uses in doing so and the robot's behaviour if an object is in this zone are adjustable.

With the integration of the GR00T foundation model and Apollo, Apptronik allows developers to teach the robot general tasks via text, video and human actions. It has long been possible to train robots by performing actions for them. Robots thereby replicate human movements.

The integration with Project GR00T's foundation model goes further and allows robots to recognise its environment based on its observations and predict which next step it needs to perform to achieve its goal.

At GTC 2024, Apptronik demonstrated, among other things, how Apollo can operate a juicer thanks to its integration with the GR00T foundation models, and then serve the juice to a person. To make this possible, Apollo is equipped with NVIDIA Jetson AGX Orin and Jetson Orin NX modules, which is hardware for running autonomous systems. Nvidia offers seven types of Jetson modules. The entry-level model is the Jetson Orin Nano, while the Jetson AGX Orin modules are the high-end model.

The Apollo robot, by the way, is not new and was launched in August 2023. Apptronik says it is targeting friendly interaction, mass production and safety with the robot. It wants to transform the way industry works, improving the human experience.

Apptronik also announced a collaboration with car manufacturer Mercedes-Benz. Together, the parties aim to identify applications for advanced robotics in Mercedes-Benz's production processes. As part of this, the carmaker will integrate Apollo into its production process.

Among other things, the parties see opportunities around automating tasks in environments designed for humans. This eliminates the need for extensive modifications in factories, and simplifies the embrace of robots. Humanoid robots can, among other things, help automate physically demanding, repetitive and boring tasks for which it is difficult to find staff.

Further information on Apollo is available here. More information on the Project GR00T foundation models is available here.

Author: Wouter Hoeffnagel
Image: Apptronik