A Galaxea representative discusses how G0.5 uses large-scale pretraining to generalize robot manipulation across objects, scenes, and instructions.
Previously, a new task often required fresh data collection, model tuning, and adaptation to a particular setting. Robots could manage familiar tasks but had difficulty with unfamiliar items, surroundings, and commands.
Galaxea presents G0.5 as a step in its robot learning work. Training across many tasks appears to give it reusable and combinable abilities, including picking up and setting down objects, moving them by pushing or pulling, and operating openable or closable items.
G0.5 can take language instructions in scenes with different layouts and objects it has not encountered before. The examples build on its basic manipulation abilities to handle changes in both the setting and the items.
Starting from the pretrained model, Galaxea says it can teach extended tasks with only a small number of examples. One example is picking up varied objects and placing them in varied locations.
Galaxea frames many practical tasks as combinations of fundamental robot abilities. It says G0.5 improves performance without task-specific setup and points toward building robot competence through pretraining rather than memorizing individual assignments.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Hang Zhao, Galaxea co-founder and CTO and Tsinghua University assistant professor, describes the company’s work on robots, datasets, and foundation models at a Galaxea sponsor talk at CoRL 2025.
