- The problem
- Robots often choose their next move from what they see right now, so they can lose track of earlier steps or repeat a motion back and forth. Meanwhile, many powerful robot models are difficult to reproduce because their training data or methods are not public.
- The idea
- The paper presents a robot-control model trained using public data and releases its full pipeline. It also gives the model a small, reusable memory: instead of reprocessing old camera images, it stores compact summaries of recent moments, like brief notes a person keeps while following a multi-step instruction.
- What it got
- The model is a leading public-data system on two robustness tests and performs competitively on other robot tasks. With memory, back-and-forth motion episodes were about four times less common.