Part 10/12:
While promising, the researchers acknowledge that current results are primarily focused on short, quick tasks (lasting seconds or less). Extending this to longer, more complex activities remains a key goal. They also aim to improve the system’s reliability, as current success rates, while impressive, still hover below optimal levels.
Another exciting avenue is enabling HPT to process unlabeled data, akin to GPT-4’s ability to understand context without explicit annotations. This would dramatically expand the available training data, further boosting the model’s versatility.