RoboReward

RoboReward Links: Paper · RoboRewardBench Leaderboard RoboReward is a dataset for training and evaluating general-purpose vision-language reward models for robotics. Each example pairs a task instruction with a real-robot rollout video and a discrete end-of-episode progress reward score in {1,…,5}. RoboReward is built from large-scale real-robot corpora including Open X-Embodiment (OXE) and RoboArena. Because OXE is success-heavy, we generate additional negatives and near-misses… See the full description on the dataset page:

Source: Hugging Face Hub (teetone/RoboReward). Metadata imported from the dataset’s Hub tags.