LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints

LIBERO-Industry. Benchmarking Robotic Manipulation under Industrial Process Constraints. By Ming Shin Cindy Sung and eleven others, from Tongji University, the Chinese University of Hong Kong, and two other institutions. Published on the fifteenth of October. Its code repository is worked out: about two hundred source files, with a readme and a licence.

Flexible manufacturing, such as high-mix, low-volume electronics assembly, needs robot policies that adapt as products and production conditions change. Those settings come with industrial process constraints: product variants, material that moves, execution windows that close, and tight geometric tolerances. Existing evaluation frameworks do not systematically capture these constraints when they assess generalist robot policies. A policy can look capable when nothing moves and still be untested against a part that is moving and a window that is closing.

LIBERO-Industry turns those constraints into controllable task conditions and measurable success criteria. It is built on LIBERO, a widely used robot learning benchmark, so the industrial conditions arrive on familiar ground.

There are three classic task scenarios. Static-to-Dynamic loading, from a fixed place onto moving flow. Dynamic-to-Static unloading, the reverse. And Dynamic-to-Dynamic transfer, from one moving flow to another. Across the suites there are two hundred and sixteen tasks. They systematically vary four things: object geometry, conveyor speed, how long the interaction window lasts, and how large the geometric acceptance region is.

Five generalist robot policies were tested, including vision-language-action models and world-action models. Performance degrades progressively as industrial constraints are introduced, and degrades further as they are combined. A second experiment used a shared fine-tuning and evaluation protocol. There, co-training with LIBERO-Industry achieved higher observed target-transfer completion than the original-LIBERO condition.

The design choice to notice is the dials. Geometry, conveyor speed, window length and acceptance region are separate controls, so a falling score can be read against the constraint that was turned up, alone or in combination. That is the measurement the abstract says earlier frameworks were not making: not whether a generalist policy can manipulate, but how its success holds up as the conditions of a real production process are layered on one by one. The co-training result adds a second use, as training material and not only as a test.