MIT researchers develop new approach for training general purpose robots
MIT develops new method for robot training using HPT inspired by GPT-4.
MIT researchers have developed the Heterogeneous Pretrained Transformers (HPT) approach, inspired by the success of models like GPT-4, to advance the training of general-purpose robots. This method combines diverse data, such as human demonstration videos and simulations, into a singular system, allowing robots to adapt to a wide array of tasks.
The use of a large dataset of 52 datasets, incorporating over 200,000 robot trajectories, allows for effective knowledge transfer and reduces the need for fine-tuning with specific data. HPT has demonstrated superior performance compared to traditional training approaches, achieving more than a 20% improvement even in tasks that significantly differ from the pretraining data.
The project received funding from the Amazon Greater Boston Tech Initiative and the Toyota Research Institute. MIT's team envisions a future where a universal 'robot brain' can be applied to various robots without specialized training, similar to developments in large language models.