Perhaps one of the hardest tasks for a robot is understanding that a person needs help, and not interfering when everything is going fine. A new benchmark at the intersection of Theory-of-Mind and embodied AI proposes to test exactly this: the agent must choose the moment to intervene on its own, rather than merely demonstrating a "theoretical" understanding of another's state.
From Questions to Action
Existing ToM benchmarks for embodied agents are most often built on a single scheme: the agent is shown a scene, then asked questions about a person's beliefs and intentions, or asked to predict the next action at the scenario level. Such tests evaluate reasoning after the fact rather than the ability to interact with a changing environment in real time. Yet it is precisely continuous perception and timely reaction that distinguish a helpful assistant from an algorithm that "answered correctly on an exam."
A Closed Loop of Precise Intervention
The new MindHelper challenge translates this task into a closed loop. The agent must constantly observe the environment, maintain separate beliefs about each person in the scene, understand when a person is genuinely struggling, and only at that moment perform a specific action. Equally important is the flip side: if intervention is not required, the agent must stay silent and do nothing.
The work builds on the previous robot-centric approach MindPower, where reasoning follows the "from perception to action" principle. However, in the new scenario this is not enough: it is necessary not only to correctly decide what to do, but also to correctly decide whether to do it at all. This design brings the benchmark closer to real-world interaction, where excessive help can be just as annoying as its absence.
How MindClaw Works
To tackle this challenge, the authors propose the MindClaw framework, which combines three components:
- Actor's belief table — a structure that stores assumptions about the goals and knowledge of a specific person at the current moment.
- Embodied cognitive skills — a set of abilities that allow the agent not only to reason but also to perform physical actions in the environment.
- Trigger-based cognitive dispatcher — a mechanism that answers the key question: has the reason for intervention occurred. The dispatcher either triggers an action or keeps the agent in observation mode.
The architecture resembles a "cognitive claw": it captures information, holds it in the form of beliefs, and releases it only when there is a real task for the robot nearby.
What the Experiments Showed
The results are telling. The MindClaw framework achieves 36.63% precise intervention and 14.36% task-solving accuracy. For comparison, direct VLM baselines that do not use the proposed architecture achieve only 12.05% and 3.80%, respectively. In other words, the gap is nearly threefold in intervention quality and nearly fourfold in task success.
This confirms that having an explicit belief model and a separate trigger for actions gives the agent a significant advantage over the "raw" use of vision-language models. What is especially valuable is that success is measured not by answer accuracy but by correct behavior in a closed loop: the robot must be equally right both when it helps and when it stays on the sidelines.
Work Status
A preprint describing the new benchmark has been published on arXiv under identifier 2606.01063. According to the authors, the paper is an extended version of the earlier MindPower work presented at CVPR 2026, and is accompanied by an additional set of experiments. The current version, marked as v3, was updated on August 24, 2026, and the project itself is still under development. So, most likely, we should expect clarifications and new results in this direction in the near future.
