SafeLIBERO Benchmark Overview

LIBERO is a widely used benchmark in the field, which comprises LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long, focusing on spatial reasoning, object manipulation, goal-conditioned tasks, and long-horizon planning, respectively. However, since objects in the original LIBERO environments rarely collide with the robot arm during movement, it is difficult to evaluate the safety capabilities of our method.

Therefore, we propose SafeLIBERO to assess performance in complex environments. Specifically, we select four tasks from each LIBERO suite and further divide each into two scenarios with different safety levels based on the degree of interference posed by the added obstacle:

  • Level I: Scenarios where the obstacle is positioned in close proximity to the target object.
  • Level II: Scenarios where the obstacle is located further away but obstructs the movement.

It is worth noting that for some tasks, the distinction between these two intervention levels may be less obvious. For each scenario, we randomize the positions of obstacles and other objects within a small range over 50 episodes. We employ a diverse set of objects as obstacles, including moka pots, storage boxes, milk cartons, wine bottles, mugs, and books. We construct 4 suites comprising 16 tasks and 32 scenarios, totaling 1600 episodes.

SafeLIBERO-Spatial
Level I
Level II
SafeLIBERO-Goal
Level I
Level II
SafeLIBERO-Object
Level I
Level II
SafeLIBERO-Long
Level I
Level II