Can a learned push transfer to a new object?

Learning from a physical simulation

An isolated experiment based on historical B29 learned a motor forward model from actual simulator stereo observations and fingertip pushes. On unseen capsule objects, directional tasks succeeded in 18 of 18 trials, compared with 9 of 18 for an untrained frozen baseline.

This is evidence of a useful difference on that test. Prediction error still remained around 6.1–6.9 centimetres, so the result is not precision placement. It also does not establish general English understanding.

When replay helps one thing and harms another

Replay improved some old-object recall but harmed novel prediction. That candidate was rejected. Retaining the trade-off prevents improved familiarity from being mistaken for improved transfer.

The work used an isolated historical branch and was not current-runtime integration. Native autobiographical learning credit remained zero. The result belongs to this experiment, with its own materials and controls.