Researchers led by Wenshuo Chen have built Trajectory-Calibrated Regression (TCR), which merges several specialised robot policies into a single one without extra gradient training, according to a paper on arXiv. It fits the merged weights on inputs collected from the specialists' own task runs and from intermediate steps of their action generation.

The merged model keeps the original structure and needs no routing between experts. On the LIBERO benchmark, TCR merged four specialist versions of the π0.5 model and succeeded in 77.83% of 1,200 test episodes. Other merging methods did 9.50 to 33.00 percentage points worse.

Source: arXiv Robotics (cs.RO)

Background

Wenshuo Chen

Robotics researcher

Robotics researcher.