Progress Reward Modeling for Robotic Learning: A Comprehensive Survey

This survey proposes a unified framework for progress reward modeling in robotic learning. It organizes prior work around three connected perspectives: the model interface, the mechanisms used to construct progress signals, and the data and benchmarks used for supervision and evaluation. The authors compare observation inputs, goal specifications, output formats, supervision sources, and evaluation protocols. The survey argues that inconsistent choices across these dimensions make results difficult to compare and obscure what each benchmark actually validates. It also summarizes current limitations and outlines future research directions.