Robots that assemble wires and cables face a harder problem than those handling rigid parts: a cable's shape is effectively infinite-dimensional and changes with every contact. Existing benchmarks for deformable linear objects (DLOs) are often tied to specific hardware or lack the fixtures found in real factories. The new WireCraft benchmark, accepted as a spotlight at CoRL 2026, aims to close that gap with configurable tasks, two complementary physics models, and trajectories collected from both simulation and a physical UR5 robot.
The benchmark spans three industrial task families—connector insertion, clip routing, and channel seating—and evaluates reinforcement learning, imitation learning, and vision-language-action models under shared metrics. The authors report that privileged state-based RL solves a representative setting in each family with over 82% success, confirming the tasks are well-posed. Image-based policies fare worse: ACT and DiT VLA both fall below the privileged State PPO baseline.
On the physical robot, the paper finds that supplementing real demonstrations with simulated data improves success by 15 to 65 percentage points in the best-performing mixture for each task. The overall message is that industrial DLO manipulation is tractable when the policy gets full state information, but remains an open challenge for vision-based learning—exactly the kind of gap a shared benchmark is designed to expose.