Unified Humanoid Get-Up Seven-Morphology Suite
MIT仓库提供Sigmaban、Wolfgang、NUGUS、Bez1/2/3和OP3-Rot七种小型人形MJCF、CrossQ训练/评测脚本与1.66MB ONNX，可在RTX3090约1.5小时训练600k步策略。它是低成本仿真训练套件，不是大规模离线轨迹库；论文无真机，六个留一法未见本体仅17–42%，Wolfgang为72±21%。
- 开放状态
- 项目制品开放·上游条款另适用
- 许可证
- MIT root license; imported morphology meshes/assets require upstream review
- 商业使用
- code permitted under MIT; asset-level use subject to source terms
- 人工评审结论
- public_mit_seven_morphology_mujoco_training_suite_with_embedded_onnx
- 机器人
- Sigmaban、Wolfgang、NUGUS、Bez1、Bez2、Bez3、OP3-Rot
- 任务
- prone fall recovery、supine fall recovery、side fall recovery、cross-morphology zero-shot evaluation
- 模态
- joint state、torso pitch、target joint action、domain-randomized dynamics、success outcome
- 格式
- MJCF/XML, Python, YAML, pickle initial states, notebook analysis and ONNX
- 采集/生成方式
- MuJoCo with Stable Baselines X / CrossQ
- 已知问题
- Pure simulation; mostly kid-size morphologies; no adult humanoid or contact-force model; LOO diagonal is 17–42% for six morphologies; ONNX lacks a model card/seed mapping; no independent replication.