MIT breakthrough could transform robot training

MIT researchers have developed Heterogeneous Pretrained Transformers (HPT), a new robot training method that combines diverse data from multiple sources. HPT outperformed traditional methods by over 20% in simulated and real-world tests. The system uses a dataset of 200,000+ robot trajectories across four categories, enabling robots to learn from various experiences. Researchers aim to create a universal robot brain applicable to any robot without additional training.