Towards Self-Evolving Embodied Foundation Model via General-Purpose In-Context Learning

 

What is In-Context Learning (ICL)?

In-Context Learning (ICL) enables AI models to understand and execute tasks by learning directly from contextual information provided during inference, without requiring task-specific fine-tuning.

What is General-Purpose In-Context Learning (GPICL)?

ICL only addresses few-shot supervised learning

GPICL learns on-the-fly but takes many steps further:

? Many-shot and Life-Long Learning

? Learning from experiences and external feedbacks

? Learning by versatile paradigms, e.g., imitation, reinforcement, unsupervised.

What are the challenges?

? Lack of scalable, high-diversity decision-making tasks

? Training to incentivize ICL and reasoning, e.g., ICRL, is difficult and not scalable

? Training with long sequences and contexts is even less efficient

z6首页OUL: The Large-Scale Meta-Training Framework

Our cutting-edge framework, z6首页OUL, powers Embodied FM with:

? Procedurally Generated Tasks with High Quality

? Decoupled Policy Distillation (DPD): As low-cost as supervised learning, as high-performances as reinforcement learning

? Linear Attention and Chunk-wise Training: easily scale to arbitrary context length

 

Codes & Papers 

airs-cuhk/airsoul: Next-gen Foundation Model for Embodied AI
Towards Large-Scale In-Context Reinforcement Learning by Meta-Training in Randomized Worlds
Putting the Smarts into Robot Bodies
Empowering Virtual Agents With Intelligent Systems
Context and Diversity Matter: The Emergence of In-Context Learning in World Models
In-Context Learning can Perform Continual Learning Like Humans

 

Demonstrations

高低文自主进化