huzican0419/embodied-mcts-habitat

Embodied MCTS — Habitat Offline Samples

Offline Monte-Carlo Tree Search (MCTS) planning trajectories collected on the EmbodiedBench EB-Habitat benchmark. 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 planners.

Files

File	Source planner	Size
habitat_offline.tar.gz	Qwen3.5-27B (open-source VLM planner)	~1.7G
habitat_offline_gpt4o.tar.gz	GPT-4o (Azure)	~1.5G

Note: navigation_offline* (EB-Nav) tars will be added in a later revision.

Each archive extracts to a directory with the same name and the following layout:

habitat_offline/
├── config.json                      # MCTS / sampling hyper-parameters
├── episode_metrics.jsonl            # one line per episode (success, #nodes, ...)
├── offline_mcts_samples.jsonl       # one line per (state, plan) training sample
└── branch_images/
    └── <eval_set>/episode_<id>/episode_<id>_step_<t>_branch_<bid>.png

offline_mcts_samples.jsonl field schema:

Field	Type	Description
instruction	str	Natural-language task instruction.
action	list	Planned action sequence (action_id, action_name).
image_path	str	Path to the observation PNG (inside branch_images).
history	list	Past actions before this state.
target_return	float	MCTS-estimated return for this plan.
mcts_q	float	Final MCTS Q-value of the chosen edge.
mcts_visit_count	int	MCTS visit count of the chosen edge.
target_confidence	float	Visit-count-based confidence in [0, 1].
planner_rank	int	Rank among siblings (1 = best).
weight	float	Suggested training weight (= visit count).
invalid / done	bool	Whether the plan was rejected / ended the episode.
env_id / task_id / subset_id / group_id	str	EmbodiedBench identifiers.

Eval-set composition

habitat_offline (Qwen3.5-27B planner): 5089 samples

• base: 233, common_sense: 862, complex_instruction: 646
• long_horizon: 1138, spatial_relationship: 1455, visual_appearance: 755

habitat_offline_gpt4o (GPT-4o planner): 4835 samples

• base: 296, common_sense: 915, complex_instruction: 780
• long_horizon: 1013, spatial_relationship: 1175, visual_appearance: 656

Download

pip install -U "huggingface_hub[cli]"

# 1. Download the tar.gz files (replace REPO_ID with this repo's id).
hf download REPO_ID \
  habitat_offline.tar.gz \
  habitat_offline_gpt4o.tar.gz \
  --repo-type dataset \
  --local-dir ./embodied_mcts_habitat

cd ./embodied_mcts_habitat
tar -xzf habitat_offline.tar.gz
tar -xzf habitat_offline_gpt4o.tar.gz

Or with Python:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="REPO_ID",
    repo_type="dataset",
    local_dir="./embodied_mcts_habitat",
    allow_patterns=["*.tar.gz", "README.md"],
)

MCTS hyper-parameters

Both splits were collected with num_simulations=8, candidate_size=16, max_plan_len=10, gamma=0.98, c_puct=8.0, temperature=0.7, seed=0, n_shots=5, env_feedback=1, down_sample_ratio=0.8. See config.json inside each archive for the full sampler config.

Citation

If you use this data, please also cite the upstream EmbodiedBench benchmark.