A new arXiv paper, titled "The Bitter Lesson Comes to Robotics," examines a recent shift in robot control. According to the abstract, general-purpose multimodal language models are increasingly capable of directly controlling robotic arms, for example by commanding end effector poses to manipulate objects. The authors say they are tracking the development of these capabilities.

The title references the well-known "bitter lesson" argument that progress in AI has historically come from scaling computation and data rather than from hand-crafted techniques. The abstract does not yet spell out the paper's conclusions, but its framing suggests the authors see a similar dynamic emerging in robotics.

Because the abstract is brief, the full scope of the work remains unclear from the announcement alone. The key claim so far is that large language models are becoming practical as direct controllers for physical manipulation, not just as planners or advisors.