README.md

metadata

license: cc-by-nc-4.0
language:
  - en
pretty_name: EmbodiedMemory-Bench
size_categories:
  - 1K<n<10K
task_categories:
  - robotics
configs:
  - config_name: default
    data_files:
      - split: all
        path: data/embodied_memory_bench.parquet
dataset_info:
  - config_name: default
    features:
      - name: id
        dtype: string
      - name: family
        dtype: string
      - name: family_id
        dtype: string
      - name: difficulty
        dtype: string
      - name: episode_name
        dtype: string
      - name: description
        dtype: string
      - name: scene
        dtype: string
      - name: probe_id
        dtype: string
      - name: probe_instruction
        dtype: string
      - name: context_sessions
        dtype: int32
      - name: context_steps
        dtype: int32
      - name: memory_cues
        dtype: string
      - name: episode_path
        dtype: string
      - name: preview_image
        dtype: string
    splits:
      - name: all
        num_examples: 2554
tags:
  - embodied-ai
  - embodied-memory
  - multimodal
  - long-horizon-planning
  - ai2-thor
  - procthor
  - benchmark

EmbodiedMemory-Bench

Benchmarking Embodied Memory for Long-Horizon Embodied Tasks

Remember what the world was. Track how it changed. Act beyond the moment.

2,554 episodes · 4 memory challenges · AI2-THOR & ProcTHOR

Benchmark

EmbodiedMemory-Bench (EMem-Bench) evaluates memory through actions in an interactive environment. An agent first observes multimodal interaction history without knowing its future task. It must then use that experience to complete a later task: remember a detail, update a changed state, avoid a previous failure, or apply a learned rule.

Memory challenge	What the agent must remember	Episodes
Passive Observation	An object’s location seen earlier, amid unrelated activity	1,036
Dynamic Tracking	The latest location or state after an object changes	1,052
Interaction Failure	A constraint revealed by an earlier interaction outcome	263
Experience Generalization	A pattern learned from corrections and applied to a new object or scene	203
Total	Four complementary memory capabilities	2,554

Interaction Failure episodes are stored under interaction_experience/. The dataset contains historical RGB observations, action and feedback traces, task probes, and evaluation annotations. Benchmark membership is fixed by the Full-2554 manifest. The Dataset Viewer uses the compact all split at data/embodied_memory_bench.parquet, with the same records available as JSONL. Each row includes paths to the complete episode record and a representative observation.

How the benchmark is constructed

Construction combines simulator-grounded scenes, task-specific memory cues, and distractor trajectories. The accompanying paper describes the execution checks and quality review used to retain the final episodes. Construction and evaluation resources will accompany the public code release.

Quick start

Download the complete episode and image package:

pip install -U huggingface_hub
hf download lzLiang/EmbodiedMemoryBench \
  --repo-type dataset \
  --local-dir EmbodiedMemoryBench-Data

Read an episode through the portable index:

import csv
import json
from pathlib import Path

root = Path("EmbodiedMemoryBench-Data")
with (root / "browse.tsv").open(encoding="utf-8") as f:
    index = list(csv.DictReader(f, delimiter="\t"))

row = index[0]
episode_path = root / row["episode_json"]
episode = json.loads(episode_path.read_text(encoding="utf-8"))
print(row["family"], episode["episode_id"])
print("Context sessions:", len(episode["sessions"]))

This release is an episode-and-media package. The index provides paths to its nested JSON records; downloading the repository also retrieves the context images. The accompanying evaluation code supports full_context and emem modes; its public release will provide simulator setup and execution instructions.

Repository layout and episode schema

EmbodiedMemoryBench-Data/
├── episodes/
│   ├── passive_observation/<episode>/
│   ├── dynamic_tracking/<episode>/
│   ├── interaction_experience/<episode>/
│   └── experience_generalization/<episode>/
├── manifests/
│   ├── full2554.jsonl
│   ├── <family>.jsonl
│   └── audit_report.json
├── browse.tsv
├── metadata.json
└── verification_report.json

Each episode directory contains its JSON record and historical images. Use browse.tsv or a family manifest’s organized_view.episode_json to locate the downloaded JSON. The canonical manifest’s source.episode_path records the current research-workspace path, and source.original_episode_path preserves its earlier location as provenance. The portable download paths are given by browse.tsv and organized_view.episode_json.

Field	Contents
episode_id	Stable episode identifier
sessions	Context trajectories with observation references, actions, and feedback
sessions[].micro_probe	Session-level probe, when present
macro_probe	Final probe, when present
memory_cues, hidden_rules	Evaluation annotations for planted cues and constraints
scene_metadata	Scene provenance and setup metadata

The raw files include evaluator annotations. Use the evaluation code to construct model inputs: future task information must stay out of history ingestion, and hidden rules, cue labels, expected actions, and privileged simulator state must stay evaluator-side.

Frozen membership and integrity records

The metadata, membership audit, and verification report record the release counts and integrity checks. The verification report records 2,554 verified episodes and no issues. Each canonical manifest row includes an episode JSON SHA-256.

Canonical manifest SHA-256:

b06f91f5609cfc1fb75696cf4abb8ec1fbd9b86131091e0fd52fd9dc38e64677

Every natural-language field in every episode JSON is now English, including nested session names, trajectory descriptions, action text, cue descriptions, probe context, hidden-rule descriptions, and evaluation explanations. All 2,554 episode JSONs are validated after decoding Unicode escapes: zero Chinese characters in any object key or string value. See the complete language validation report. Episode membership, simulator identifiers, numeric values, image files, and evaluation structures are preserved; textual annotations and integrity hashes have been updated. The original-language version remains available at revision 28694bf83b78049ac96ba01c9ffb13aeac7eaef9.

The release is the complete evaluation set; manifest rows use split: all. No train/validation partition is declared here.

Evaluation

The benchmark measures Success Rate (SR ↑) and Error Recurrence Rate (ERR ↓). Report per-family scores and their equal-family macro average so the smaller task families remain represented.

Sources and license

The benchmark is released under CC BY-NC 4.0. It builds on AI2-THOR and ProcTHOR; their environments, assets, and dependencies remain subject to their respective upstream terms.

Reference

EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks

Read the paper: arXiv:2609.28236.

@misc{embodiedmemorybench,
  title         = {EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks},
  author        = {Liang, Lizhou and Zhong, Xinyu and Pan, Miao and Zhou, Xiaohe and Liu, Xuanyu and Li, Qinfeng and Li, Peng and Chen, Jintao and Zhang, Xuhong and Zhang, Wenqi},
  year          = {2026},
  eprint        = {2609.28236},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.28236}
}

Project resources: project page · code.