README.md

metadata

license: gemma
library_name: transformers
base_model:
  - Dexmal/DM05
tags:
  - robotics
  - vision-language-action
  - dm05
  - opendm

DM05-mem

OpenDM-format BF16 checkpoint for DM05-mem (32-slot history EEF). Use with OpenDM playground/dm05_mem.py. See the DM05 Inference Guide.

Weights: BF16 model.safetensors.

Model Card

Field	Value
Config	playground/dm05_mem.py
Env vars	DM05_MEM_CHECKPOINT
OpenDM robot_type	ARX5 (playground default)
Control	EEF (control_mode=eef)
Cameras	Head / Left wrist / Right wrist
Native state / action	7 / 7 (xyz+rpy in, xyz+axis-angle+gripper out)
Defaults	action_horizon=50, is_history=true, speed=0.1, model_max_length=2048

Use with OpenDM Inference

# From the OpenDM repository root.
hf download Dexmal/DM05-MEM \
  --local-dir ./checkpoints/DM05-MEM
pip install -e ".[fast-infer]"

script/dm05_launcher.sh \
  --exp playground/dm05_mem.py \
  --task inference \
  --model-config.model-name-or-path ./checkpoints/DM05-MEM \
  --inference-config.port 7891

Override with the env var used by the playground:

export DM05_MEM_CHECKPOINT=/path/to/DM05-MEM
python playground/dm05_mem.py
python playground/dm05_mem.py --inference-config.backend fast

Runtime Profile

From playground/dm05_mem.py:

• control_mode=eef, action_mode=absolute, speed=0.1
• is_history=true, max_history_images=32
• attn: llm/action sdpa, vision flash_attention_2
• fast: prefix 2048

Files

.
├── config.json
├── model.safetensors
├── norm_stats.json
├── tokenizer.json
├── tokenizer_config.json
├── processor_config.json
├── chat_template.jinja
├── generation_config.json
└── README.md

Citation

@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}
}