Machine learningML & AI Methods

Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers

The paper presents Heterogeneous Pre-trained Transformers (HPT), a method for training robotic models across various tasks, improving the performance of fine-tuned policies by over 20% on unseen tasks in both simulated and real-world environments.