Researchers introduced the Constraint‑Referenced Strategy (CRS) to evaluate robot grasping and pushing actions. CRS looks at how the robot’s commanded motion line up with the actual motion the object’s joint can perform. It separates useful progress from wasted movement. The method only needs a model of the joint and a simple controller – no demonstrations or learning are required.

When applied to simulations of doors, drawers and faucets, CRS spotted three kinds of failures that a simple “task completed or not” check ignored. It identified reverse‑pushing, misaiming and off‑axis mistakes. In a dramatic example, a policy trained to open doors was able to pass drawer tests, yet the command it issued could not actually drive the drawer joint. The goal‑success bit said the task was done, while CRS showed the command was physically impossible.

Even if the joint axis is estimated incorrectly by up to 30 °, or the threshold used to judge the motion is off by about 8°, CRS still gives the same verdicts. The approach was tested in simulation on three mechanisms, and it focuses on the command itself, not the physical cost of the wasted motion.

CRS offers a clearer picture of why a robot fails, helping developers to refine policies that real robots can actually execute.