Recursive Video In-Context Learning for Agentic Robot
arXiv cs.AI · 6 Oct 2026, 03:59 AEST
A new method called Recursive Video In-Context Learning (RV-ICL) enhances the performance of LLM agents in robotic tasks. This approach allows agents to navigate a hierarchy of demonstration videos, improving task execution efficiency and success rates.
Key facts
- LLM agents that orchestrate frozen vision-language-action (VLA) policies improve across episodes through text memory, which records what the agent did but not how the task is done.
- We introduce Recursive Video In-Context Learning (RV-ICL), a training-free method that turns a demonstration into a hierarchy the agent navigates rather than a prompt it receives.
- The hierarchy is built from the sub-events of the demonstration, such as grasps and releases.
- Built on RPent, RV-ICL raises success from 92.6% to 96.5% on LIBERO-PRO and from 86.7% to 95.8% on LIBERO-Plus.
Read aloud
A new method called Recursive Video In-Context Learning RV-ICL enhances the performance of L L M agents in robotic tasks. This approach allows agents to navigate a hierarchy of demonstration videos, improving task execution efficiency and success rates. L L M agents that orchestrate frozen vision-language-action VLA policies improve across episodes through text memory, which records what the agent did but not how the task is done. We introduce Recursive Video In-Context Learning RV-ICL, a training-free method that turns a demonstration into a hierarchy the agent navigates rather than a prompt it receives. The hierarchy is built from the sub-events of the demonstration, such as grasps and releases. Built on RPent, RV-ICL raises success from ninety two point six percent to ninety six point five percent on LIBERO-PRO and from eighty six point seven percent to ninety five point eight percent on LIBERO-Plus.
Sources in this cluster
- Recursive Video In-Context Learning for Agentic Robot
arXiv cs.AI · 6 Oct 2026, 03:59 AEST