Apptronik Working to Integrate Apollo With NVIDIA’s Robot Foundation Model Project GR00T

The Robot Report reported on March 23, 2024, that Apptronik was working to integrate its humanoid robot Apollo with NVIDIA’s general-purpose robot learning foundation model Project GR00T. The article also said NVIDIA CEO Jensen Huang demonstrated Apollo operating a juicer and serving juice during a keynote.

The plan described by The Robot Report envisions using text, video and human demonstrations as task inputs and translating them into Apollo’s actions. However, the demonstration featured in the article was an example of a specific task, and no figures confirming reproducibility or safety in real-world operations were available. The model’s integration and its deployment in the field therefore need to be considered separately.

The technical focus of this report is not the development of the humanoid itself, but how to connect inputs to a foundation model with a robot’s actions. According to the article, the concept is to treat text, video and human demonstrations as task instructions and translate them into actions by Apollo. However, the report describes the companies as working on the integration; it does not show that it has been completed or that performance reproducible in general environments has been confirmed. The juicer operation shown in the keynote was also presented as a demonstration, with no details on task success rates, conditions or behavior over repeated runs. This approach needs to be evaluated by distinguishing between a model’s ability to process instructions and a robot’s ability to act safely in response to its environment. The article reports Apptronik’s explanation that Apollo perceives its environment and predicts its next action, but does not specify how perception works, how actions are generated or what range of data is used for training. It is therefore important not to equate the announced integration plans with capabilities verified on a physical robot. A connection to later introductions of robotics technologies can be found in NVIDIA Introduces “SOMA Retargeter” to Adapt Human Movements for Robots. The information provided in that article says human-motion demonstrations are used as a starting point for teaching robots to move. This overlaps with the approach to inputs described in the report on Apollo and Project GR00T, but no information shows that the two are technically linked or share an implementation. NEXCOM and NexCOBOT Reportedly to Showcase RoboWIZ and Four Technology Partners at GTC Berlin, which covers NVIDIA IGX Thor, is also a separate effort involving computing platforms for robotics; no direct connection to this case has been confirmed. The next points to verify are when the integration will be completed, the range of tasks it can perform, its reproducibility across different environments, and how it will stop and ensure safety when failures occur. These details are not included in the report provided, limiting the basis for assessing the technology’s practicality or scope of deployment at this stage.