Offline Monte-Carlo Tree Search (MCTS) planning trajectories collected on the
EmbodiedBench EB-Navigation
benchmark (point-goal / object-goal navigation in Habitat scenes).
Each sample is a (instruction, observation_image, history) -> plan tuple
annotated with the MCTS Q-value used as a target return, suitable for offline
imitation / Q-learning / preference training of VLA navigation policies.
File
Source planner… See the full description on the dataset page:
https://huggingface.co/datasets/huzican0419/embodied-mcts-navigation.
