What they found
The authors built Mind2Act World, a simulated benchmark of 30 manipulation tasks from five real-world application domains. Each task has three difficulty levels and two scene settings (Clean and Random), giving 180 configurations. Tasks are split into three views: Mind (decisions based on past observations), Act (precise and timely execution), and Mind2Act (closed-loop tasks where action results guide later decisions). They tested multimodal agents, vision-language-action models (VLAs) and world-action models (WAMs). Skill in one view did not carry over to Mind2Act. The strongest agent succeeded in 52.7% of Mind episodes and 39.0% of Act episodes, but only 12.0% of Mind2Act episodes. The authors say this agent keeps track of the task but struggles with sustained contact. End-to-end policies carry out practiced motions but keep little task state.
Why it matters
Many robot benchmarks mix reasoning and execution, so a failure is hard to explain. This benchmark separates the two and tests how they combine. That helps builders see whether a system lacks a skill or fails to combine skills it already has. The authors conclude that combining reasoning and execution through environmental feedback is the main open problem.
How it changes the race
If the results hold, labs building embodied SI may shift effort from improving reasoning or motion on their own to linking them through feedback. The benchmark gives a way to find which part is failing. It is a measurement tool, not a new capability.
The abstract, by the authors
Embodied intelligence aims to let robots perform complex physical tasks with human-like adaptability and reliability, which requires three complementary capabilities: history-based cognitive reasoning, precise and timely physical execution, and the composition of reasoning and execution through environmental feedback. Yet existing benchmarks either evaluate reasoning and execution within the same tasks or assess them separately without their composition, so they reveal little about whether a failure comes from reasoning, execution, or their composition. We introduce Mind2Act World, a diagnostic benchmark of 30 simulated manipulation tasks from five real-world application domains. Each task has three difficulty levels and two scene settings (Clean and Random), yielding 180 configurations. The tasks form three views: Mind tests decisions based on past observations, Act tests precise and timely execution, and Mind2Act combines both in closed-loop tasks where action outcomes guide later decisions. Comparing the three views shows whether a system lacks a capability or fails to compose capabilities it already has. A diagnosis harness further attributes the failed episodes of three representative systems to Mind, Act, or their composition and summarizes each system in a profile card. Our evaluation of multimodal agents, vision-language-action models (VLAs), and world-action models (WAMs) shows that competence in a single view does not carry over to Mind2Act tasks. Even the strongest agent succeeds in 52.7% of Mind and 39.0% of Act episodes but in only 12.0% of Mind2Act episodes. The strongest agent retains task state but struggles with sustained contact, whereas end-to-end policies execute practiced motions but retain little task state. These findings point to composing reasoning and execution through environmental feedback as the main open problem.
Quoted from arXiv:2611.26444. The research and its words belong to its authors.
