Generalist AI Releases GEN-1.5 Robot Foundation Model
Generalist AI unveils GEN-1.5, a multimodal robot foundation model that learns new tasks instantly from a single 3-12 second demonstration.
Stock photo for illustration only, not from the actual event
- Generalist AI announces the release of the GEN-1.5 robot foundation model.
- Learns new tasks instantly from a 3-12 second demonstration.
- Processes video, sensor, language, and proprioceptive inputs.
- Currently available exclusively through direct partnerships.
Generalist AI has announced the release of GEN-1.5, a large multimodal robot foundation model designed to process video, sensor, language, and proprioceptive inputs. The model maintains a 30-second context memory and outputs action trajectories at 100 Hz.
The system has undergone continuous pretraining for over eight months using physical interaction data captured across homes, warehouses, and factories. The core mechanism relies on physical prompting, where sensorimotor examples are inserted into the context window.
Stock photo for illustration only, not from the actual event
Stock photo for illustration only, not from the actual event
The model's ability to perform tasks immediately without gradient steps highlights how in-context learning capabilities can emerge naturally from pretraining scale, mirroring how one-shot prompting first appeared in large language models like GPT-3.
Across 10 diverse tasks, one-shot in-context prompting achieved an average success rate of 59% without any prior training. Performing 10 gradient steps on five minutes of data raised that success rate to 83%, demonstrating significant efficiency in test-time training.
At present, GEN-1.5 is introduced as a research release with no public weights, API, or self-serve product available. Interested parties must establish a direct partnership with Generalist AI to gain access.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment