dm05

DM0.5 VLA with a memory of past frames (Dexmal DM05-MEM-Robodojo-Sim), bf16, through XPolicyLab’s OpenDM adapter

• Kind: policy
• Protocol: xpolicylab
• Default port: 19009
• VRAM: 13 GB
• Load time: about 240 s
• Skills that use a policy: vla
• Configurations that name it: –

What the model is told

DM0.5 VLA (Dexmal DM05-MEM-Robodojo-Sim, with a memory of past camera frames) trained on whole RoboDojo tasks: demonstrations from the start pose, with each task own instruction (joint actions, both arms)

Weights

harness fetch dm05 --deploy <machine>

Directory	Source	From
<weights>/dm05	modelscope_model	Dexmal/DM05-MEM-Robodojo-Sim

Notes

OpenDM needs torch 2.11 and transformers 5.3: give it its own interpreter with the deploy file’s python for this service.

Manifest

kind = "policy"
summary = "DM0.5 VLA with a memory of past frames (Dexmal DM05-MEM-Robodojo-Sim), bf16, through XPolicyLab's OpenDM adapter"
protocol = "xpolicylab"
port = 19009
# bf16: the released FP32 weights alone take 23.3 GB.
vram_gb = 13.0
load_seconds = 240
env = { PYTHONDONTWRITEBYTECODE = "1" }
client = { url = "{url}", scoped = true, description = "DM0.5 VLA (Dexmal DM05-MEM-Robodojo-Sim, with a memory of past camera frames) trained on whole RoboDojo tasks: demonstrations from the start pose, with each task own instruction (joint actions, both arms)" }
needs_files = ["{weights}/dm05/model.safetensors", "{xpolicylab}/policy/OpenDM"]
notes = "OpenDM needs torch 2.11 and transformers 5.3: give it its own interpreter with the deploy file's python for this service."

[xpolicylab]
policy = "OpenDM"
ckpt_name = "DM05-MEM-Robodojo-Sim"
action_type = "joint"

[xpolicylab.overrides]
model_path = "{weights}/dm05"
norm_stats_path = "{weights}/dm05/norm_stats.json"
bf16 = "true"

[[weights]]
dir = "dm05"
modelscope_model = "Dexmal/DM05-MEM-Robodojo-Sim"