Research topics in robot learning.
Start with the problem you study. Follow the connections between robot learning, contextual evidence and physical action.
Vision-language-action models
How do pretrained policies use new task evidence? Connect VLA inference, multimodal prompting and deployment adaptation, and distinguish the effect of teaching from the competence acquired in training.
Learning from demonstration
What can a robot learn from one example? Compare one-shot and few-shot imitation, human-video teaching and geometric transfer through the information each demonstration provides.
World models for robot control
When does predicting the future help a robot act? Trace the roles of video prediction, world-action models and planning, from a context-conditioned forecast to executable motion.
Memory for embodied agents
Which experience should an agent keep? Compare episodic history, retrieved guidance and reusable skills by the decisions they change and the conditions under which they remain useful.
Robot self-improvement
Does experience make the next task easier to learn? Examine active coaching, failure-driven repair and physical recursive self-improvement through what is updated, retained and tested.
A shared framework for these questions.
The Robot ICL Survey follows new evidence into execution: the actions, motion references, predicted consequences, or skill requests that a robot consumes. This connects research topics that use different models but face related questions about teaching, transfer and retained experience.
Use the four-interface taxonomy to compare these computational routes, and the evaluation framework to design tests of context use. The comparison with related surveys places this perspective alongside learning from demonstration, VLA architectures and agentic robotics.