license: cc-by-4.0
task_categories:
- robotics
tags:
- HuMeGo
- humanoid
- unitree-g1
- memory
- loco-manipulation
- benchmark
pretty_name: HuMeGo-Sim-TestScenes
HuMeGo-Sim TestScenes
Part of HuMeGo, the humanoid memory benchmark of Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation (arXiv · GitHub · project page · all HuMeGo datasets).
Test scenes for the six simulated HuMeGo tasks. Each task folder holds 90 scenes: 30 per randomization level (L0: visual and distractor randomization; L1: plus lighting; L2: plus spatial perturbation) and 30 per answer. Test seeds are disjoint from the training demonstrations and were accepted by the same scripted expert as the training data.
| Folder | Scenes | Training data |
|---|---|---|
FlashPad/ |
90 | HuMeGo-Sim-FlashPad |
HideMug/ |
90 | HuMeGo-Sim-HideMug |
ShellGame/ |
90 | HuMeGo-Sim-ShellGame |
LitLamp/ |
90 | HuMeGo-Sim-LitLamp |
BlinkPad/ |
90 | HuMeGo-Sim-BlinkPad |
TouchPad/ |
90 | HuMeGo-Sim-TouchPad |
Each folder is in LeRobot v2.1 layout with one frame per episode: the initial robot state of the scene. The scene itself is in
environment_config of meta/episodes.jsonl: task id, scene seed, answer, randomization level and the domain-randomization state
for the SIMPLE simulator. usd_path values are relative to the SIMPLE
repository root. TouchPad stores its scene under the bp2t key (answer = blink_pad, level = touch-L0/L1/L2) and adds
start_frames/, a rendering of each scene's initial state for reference. A trial succeeds when the robot places the object on
(or, in TouchPad, touches) the pad given by the answer.
License
Released under CC BY 4.0.
Citation
@misc{chen2026memorybeforegoal,
title={Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation},
author={Chenyuan Chen and Ran Ding and Lechen Zhang and Zewen He and Yoshihiko Nakamura},
year={2026},
eprint={2610.31166},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2610.31166},
}