Submitted on October 22, 2026, the AstroBench preprint studies unsupported-object capture with a fixed-base arm in zero-gravity simulation. Frozen π0.5 DROID captures 2/200 static and 4/200 moving targets. Separately cube-trained ACT reports 8/10 and 6/10. Training and detectors differ, so these figures do not rank architectures. Success in these experiments is an acquisition event, not guaranteed retention through the end. This is not a flight test.
Why an unsupported object changes the task
A table constrains some of an object's movement. For a free-floating object, attempting contact can itself change position and orientation. An attractive image of an approaching gripper therefore cannot establish whether the operation will finish successfully. A time sequence is needed to show what happens after first contact.
The editorial significance is the explicit skill-transfer problem. A solution that worked in another environment needs fresh evaluation under different physical conditions. Few successful attempts do not, by themselves, diagnose the cause. Observation, control, interface compatibility or task conditions could be involved. The source does not justify presenting those possibilities as established explanations.
How we would structure a comparison
For an owned comparative study, we would choose the same objects, starting positions and success rule. We would then hold execution time and observation delivery constant. This is a proposed plan, not an additional SIExpo measurement. The aim is to connect a difference in outcomes to the change being investigated, rather than to unequal task difficulty.
A separate measure would track retention. An object briefly between the fingers can subsequently slip away. For a useful operation, we would record both the first valid acquisition and the final state. That prevents brief contact and sustained completion from receiving the same practical interpretation.
Small studies should preserve every attempt, including failures and interruptions. Repeating one starting position is not equivalent to testing new objects. The team should define in advance which generalization it wants to measure: new motion, a new shape or a new viewpoint.
We would also preserve the scoring configuration with each result. A later revision of a detector might be reasonable, but its outputs should remain distinguishable from earlier measurements. Otherwise, an apparent improvement could reflect a changed definition of success instead of better manipulation.
A released resource and demonstrated behavior
The project has a public repository for tasks and resources. Availability enables inspection of the specification; it does not constitute independent reproduction of results. SIExpo has not installed or executed the environment. Likewise, a large expansion inventory does not mean a policy has been evaluated on every item it contains.
AstroBench is useful as a concretely defined research problem and a warning against invalid comparisons. For practical application, the next step should be traceable evaluation of retention and transfer under comparable conditions. The text remains a preprint, and that scope should be preserved whenever its results are presented.
*Cover: conceptual AI illustration, not a documentary photograph.*
*Editorial verification: 2026-10-25T07:09:22.750908+00:00.*
