Reach the target, avoid the obstacle. It is the standard test problem for robot arm control and has been for years, which is precisely why it is worth being suspicious of: a benchmark everybody has optimised against stops measuring what it originally did.

The authors point out that the tabletop version is simplified and restricted, and that whether deep reinforcement learning handles genuinely complex reach-avoid scenarios remains uncertain. Success on the easy case has been read as evidence about the hard case without anyone checking the inference holds.

Their contribution is a comprehensive benchmark for the harder setting, built on vectorized simulation. The engineering choice is what makes the science possible: running many environments in parallel is what turns an experiment that would take weeks into one that can actually be run across enough conditions to mean something.