# A2-MEAS reports results across robot simulation and household tasks

> The arXiv paper introduces a multi-agent robot-learning system and reports strong benchmark results, including improved recovery after failures in coffee preparation.

Oossa · 2026-11-09 · https://oossa.com/en/a2-meas-reports-results-across-robot-simulation-and-household-tasks

Researchers introduce A2-MEAS, a robot-learning system that uses written “correction codes” to guide actions and recover from mistakes. Its A2-WLA action model combines motion history with a model of how the world changes.

The paper reports 45.40% task success across 100 public BEHAVIOR 2026 tasks and 57.92% average success across 12 real-robot household tasks. In coffee-preparation tests, a history-conditioned adapter improved completion after failures at two filter-handling steps; the authors say they plan to release the system and training materials.

## The facts

- A2-MEAS scored 0.6411 on the 100-task BEHAVIOR 2026 public set, according to the paper.
- For two coffee filter-handling steps, post-failure completion rose from 14% to 94% and from 8% to 82%, compared with A2-WLA retries.

## Why it matters

For robotics researchers, the results suggest that recording past actions and corrections may help robots recover from household-task failures, though the paper’s reported tests do not establish how well the approach will work in other settings.

## Sources & references

1. [Toward the Era of Agentic Embodied Intelligence](https://arxiv.org/abs/2611.09124) – arXiv cs.RO, 2026-11-09

Last updated: 2026-11-09