The engineering question is how execution works and which constraints were actually tested. A diffusion-based alignment policy is combined with a contact classifier and a force-controlled unfastening controller. The initial engagement stage remains unreliable across variations; angled fasteners were not tested.

Removing a screw looks simple until a robot has to do it. The screwdriver must reach the correct point, align with the head, engage the slot and maintain contact while turning. Hemant Kumar and colleagues presented DAWN on October 29, 2026, a research system that breaks this sequence into manageable parts. Their motivation is automated disassembly, including the recovery of components from used products. A robot that merely detects a screw in an image has not yet solved the mechanical problem of getting the tool to work.

The first component uses a diffusion-based action policy, a learned method that progressively constructs movements from observations, to position and insert the screwdriver. A second model combines visual and force information to decide whether the tip has engaged securely. A third controller handles turning and removal, relying mainly on force feedback rather than slow image-based action generation. The authors impose a reaction-force limit of 20 newtons during this stage. The separation matters because positioning, recognizing contact and applying torque have different sensing and timing requirements.

The experiments used a real Franka Emika robot arm fitted with an electric screwdriver. The researchers collected 199 expert demonstrations for the alignment stage and 6,000 contact images for training the engagement classifier. In a reported comparison, the alignment-and-engagement policy succeeded in roughly half the trials, comparable to trained human teleoperators under that test setup. The contact classifier reached an accuracy of 0.90. Once the tool was properly engaged, the tested removal stage performed much more consistently, showing where the difficult bottleneck actually lies.

This is not evidence that a robot can already dismantle arbitrary electronics without supervision. Screw heads with unfamiliar appearances, changes in lighting and small positioning errors can defeat the initial stage. The researchers did not evaluate non-parallel screw orientations, and they identify sensitivity to variation in the paper. A production system would also need to detect screws on complex assemblies, recover from slips and verify that the surrounding components remain undamaged. DAWN offers a useful lesson: the hardest part of a seemingly simple task may be establishing the right contact before the visible action begins.

The useful question is which behavior was actually tested and which remains a research ambition. A successful demonstration is exciting, but understanding the technology also means asking what happened during failures and which test conditions stayed fixed.