近期参加智源人工智能大会,其中在听到USCD助理教授苏昊的分享时,感觉深受启发,因此将其对于具身智能的几个观点总结如下,供大家学习分享。
几个认知:
1、Perception,cognition and action are intimately coupled,and form a closed loop.
感知、认知和行动是密切相关的,并且将构成一个闭环
2、Intelligence emerges in the interaction of an agent with an environment and as a result of sensorimotor activity.
智能在智能体与环境的交互中涌现,是感觉运动行为的结果。
3、Key question of Embodied AI: emergence of concepts and representation learning by coupling perception,cognition,and action.
具身智能的核心问题:通过耦合感知、认知和行动,来达到概念涌现和表征特征学习。
数据如何获得(how are data acquired):
互联网智能时代,人类制作数据集,人类去做标注,用算法建立映射。
具身智能时代,一个机器人应该能自主的去收集,应该能主动的跟环境交互来获取数据,数据收集人不止是人,更是机器人自身,可以通过历史数据来主动学习。-> 这会涉及到决策论中的一个本质矛盾:探索和利用。
数据如何处理 (how to process data):
数据从感知端流动到决策端,中间会经过一次对世界的建模。
任务驱动的表征学习 Task-driven repr. learning
物体功能的理解 Functional understanding
概念的涌现认知 Emergence of new concepts
如何评估数据性能(How to Evaluate the performance?)
Challenge 1: variation of object arrangement and plan horizon
Challenge 2: complex skills with robustness to variations
Tips:
Task completion rate
Sample complexity (amount of interactions)
Compositional generalizability 决策是长sequence,需要很强的组合泛化能力
具身智能的核心挑战:技能学习(Key challenge of embodied Ai : Skill learning)
Caveat: skill chaining 短距技能
The full task is solved by skill chaining.
Acquiring Robust Basic skills is hard,基础操作技能是很大的瓶颈
网友评论