A new paper on arXiv, titled "Vision as Proprioception: Estimating Humanoid Robot Pose from Video," takes on a problem the authors say has not been explored: determining a humanoid robot's internal state from video alone. The abstract frames this as a practical challenge with real-world applications, though it does not yet describe the proposed method or results.
The motivation is straightforward: during interactions with a robot, its onboard state information may be unavailable due to privacy or other reasons. In such cases, video could serve as a substitute for proprioceptive sensors, allowing observers to infer the robot's pose without direct access to its internal systems.
The abstract is brief and does not detail the technical approach or performance. As a new arXiv announcement, the full text would be needed to assess the method's feasibility and accuracy. The paper is currently available on arXiv under the identifier 2610.25288.