After decades of single purpose systems, the jump to more generalized systems will be a big one. The push is towards creating robotic intelligence that can fully use the wide range of movements enabled by humanoid design.
Training general purpose robotics remains a challenge due to fragmented approaches. Future solutions may involve combining methods like reinforcement learning and imitation learning, augmented by generative AI models.
The MIT team introduced Policy Composition (PoCo) for collating information from task-specific datasets. The use of diffusion models improved task performance by 20%, enabling the robot to handle multiple tasks and unfamiliar situations.
[
Collection
]
[
|
...
]