Researchers led by Jiachen Li have examined where the gains come from when robot policies are pretrained on human data, according to a paper on arXiv. Gains credited to a data recipe may also come from the size of the subset, from variation between training runs or from later adaptation.

Using an HPT implementation on 20 MetaWorld simulation tasks, a random 5,000-recording subset matched most of the gain of a subset picked by task over the full pool of 40,000 recordings, 3.69 of 4.29 percentage points. Between subsets of the same size, the gaps were small, with 0.60 points for subsets of 5,000 recordings and 0.45 points for subsets of 1,250.

Two experiments meant to explain the gains did not show the expected effects. The authors say controls for subset size and capacity, results reported per seed and targeted tests change how apparent transfer gains should be read.

Source: arXiv Robotics (cs.RO)

Background

Jiachen Li

Robotics researcher

Robotics researcher.