Researchers from Hugging Face have published a comprehensive survey on progress reward modeling for robotic learning, aiming to unify a fragmented field. The paper, titled 'Progress Reward Modeling for Robotic Learning: A Comprehensive Survey,' addresses the challenge of providing intermediate feedback during robot task execution, beyond simple success/failure signals.
The survey organizes the field into three connected steps: the interface of a progress model (what information it receives and what signal it produces), the internal methods for constructing the progress signal, and the data and benchmarks used for evaluation. This structure connects what a progress model is, how it is built, and how its quality is validated.
The authors highlight that current methods use diverse observations, goal specifications, output signals, supervision sources, and evaluation protocols, making comparisons difficult. The survey summarizes limitations and suggests future research directions.
Additionally, the team points to their earlier work, ProgressLM (ACL 2026 Oral), a reasoning model that explicitly scores progress in long-horizon robotic manipulation tasks. They have also released a GitHub repository curating papers, code, and benchmarks related to robotic progress modeling, inviting community contributions.