← The frontier
Technology & AIAug 17, 2026

Don't Drop the BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory

Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task.

Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task. Vision-language-action (VLA) models increasingly master the individual skills, yet the chain still fails: errors compound beyond the policy's ability to correct, and one subtask silently…

The frontier is open to all. Sign in to learn this from first principles and save it to your knowledge base.