Robotics
Transformers
Safetensors
dm05
text-generation
vision-language-action
opendm
robochallenge
aloha
Instructions to use Dexmal/DM05-Table30v2-ALOHA with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Dexmal/DM05-Table30v2-ALOHA with Transformers:
# Load model directly from transformers import AutoModelForSeq2SeqLM model = AutoModelForSeq2SeqLM.from_pretrained("Dexmal/DM05-Table30v2-ALOHA", device_map="auto") - Notebooks
- Google Colab
- Kaggle
metadata
license: gemma
library_name: transformers
base_model:
- Dexmal/DM05
tags:
- robotics
- vision-language-action
- dm05
- opendm
- robochallenge
- aloha
DM05-ALOHA(RoboChallenge)
面向 RoboChallenge Table 30 v2 ALOHA 机型的 OpenDM 格式 BF16 权重。
请配合 Dexbotic-RoboChallengeInference
(configs/generalist/aloha.yaml)使用。
权重文件为 BF16 model.safetensors。
模型信息
| 字段 | 值 |
|---|---|
| 配置名 | generalist/aloha |
| 环境变量 | ALOHA_CHECKPOINT / ALOHA_NORM_STATS |
OpenDM robot_type |
Aloha |
| 控制方式 | Joint 相对动作 |
| 相机顺序 | Head / Left wrist / Right wrist |
| 平台相机映射 | cam_high → image_0,cam_left_wrist → image_1,cam_right_wrist → image_2 |
| 状态 / action 统计维度 | 14 / 14(norm_stats.json 仅含 action 分位数统计) |
| 默认推理参数 | action_horizon=25,is_history=false,inference_seed=0 |
Table30 任务
lint_roller_remove_dirt、pack_the_items、pack_the_toothbrush_holder、paint_jam、
put_the_books_back、put_the_pencil_case_into_the_schoolbag、
scoop_with_a_small_spoon、stamp_positioning、wipe_the_blackboard、
wrap_with_a_soft_cloth
分任务 horizon 见 configs/generalist/aloha.yaml 的 task_overrides
(例如 put_the_books_back=15、stamp_positioning=45、wipe_the_blackboard=15)。
使用 Dexbotic-RoboChallengeInference
cd /path/to/Dexbotic-RoboChallengeInference
export OPENDM_ROOT=/path/to/opendm
export ALOHA_CHECKPOINT=/path/to/DM05-ALOHA
export ALOHA_NORM_STATS=$ALOHA_CHECKPOINT/norm_stats.json
pip install -e "$OPENDM_ROOT"
pip install -r requirements.txt
python execute.py --config-name generalist/aloha \
user_id=YOUR_USER_ID \
submission_id=YOUR_SUBMISSION_ID
不用环境变量时也可直接覆盖:
python execute.py --config-name generalist/aloha \
checkpoint=/path/to/DM05-ALOHA \
norm_stats=/path/to/DM05-ALOHA/norm_stats.json \
user_id=YOUR_USER_ID \
submission_id=YOUR_SUBMISSION_ID
若未设置 ALOHA_NORM_STATS,客户端会回退到 $ALOHA_CHECKPOINT/norm_stats.json。
运行时配置
来自 configs/default.yaml → robot_profiles.aloha:
action_type=joint,action_mode=relativeis_history=false,add_state=true,speed=0.5inference_seed=0,resize_name=padding
目录结构
.
├── config.json
├── model.safetensors
├── norm_stats.json
├── tokenizer.json
├── tokenizer_config.json
├── processor_config.json
├── chat_template.jinja
├── generation_config.json
├── README.md
└── README_zh.md
引用
@misc{dm05,
title = {{DM0.5}: An Open-World Foundation Model for General-Purpose Embodied Intelligence},
author = {{Dexmal Team}},
month = {July},
year = {2026},
url = {https://www.dexmal.com/blog/dm0.5/index_en.html}
}