Reaction
EgoLAP links human motion to robot actions through language-based reasoning
The EgoLAP team reports 80.1% mean task progress across five behaviors on a robot setup unseen during training.
TLDR
The EgoLAP team describes a model trained on roughly 10,000 hours of human and robot trajectories. It uses language to describe actions and why a motion fits a task, helping connect different ways humans and robots achieve the same result. The team reports 80.1% mean task progress across five real-world behaviors on an unseen robot setup and 38.3% success on 14 simulated tasks, without training on simulation data.
Combined views
—
1 Source, first seen ago
