Dexmal/DM05-Table30v2-UR5

DM05-UR5 (RoboChallenge)

OpenDM-format BF16 checkpoint for UR5 on RoboChallenge Table 30 v2. Use with OpenDM third_party/robochallenge_inference (configs/generalist/ur5.yaml). See the DM05 RoboChallenge Table 30 v2 Inference Guide.

Weights: BF16 model.safetensors.

Model Card

Field	Value
Config	generalist/ur5
Env vars	UR5_CHECKPOINT / UR5_NORM_STATS
OpenDM robot_type	UR5
Control	EEF relative (single_arm_target=eef)
Cameras	Head / Left wrist
Platform cams	cam_global → image_0, cam_arm → image_1
Native state / action stats	7 / 7 (action-only quantile norm_stats.json)
Defaults	action_horizon=25, action_playback_target_steps=0, is_history=false, ur5_anchor_pitch_zero=true

Table30 Tasks

arrange_fruits, item_classification, shred_paper

From third_party/robochallenge_inference/configs/generalist/ur5.yaml:

Task	Extra runtime
arrange_fruits	ur5_anchor_roll_pitch_zero=true (roll→±π, pitch→0)
item_classification	same
shred_paper	pitch-only (profile default; no roll lock)

Use with OpenDM RoboChallenge Inference

The RoboChallenge client now lives in OpenDM at third_party/robochallenge_inference (configs/generalist/ur5.yaml). See the DM05 RoboChallenge Table 30 v2 Inference Guide.

# From the OpenDM repository root.
export OPENDM_ROOT=/path/to/opendm
pip install -e ".[fast-infer]"

cd third_party/robochallenge_inference
export UR5_CHECKPOINT=/path/to/DM05-UR5
export UR5_NORM_STATS=${UR5_CHECKPOINT}/norm_stats.json
pip install -r requirements.txt

python execute.py --config-name generalist/ur5 \
  user_id=YOUR_USER_ID \
  submission_id=YOUR_SUBMISSION_ID

Override without env vars:

python execute.py --config-name generalist/ur5 \
  checkpoint=/path/to/DM05-UR5 \
  norm_stats=/path/to/DM05-UR5/norm_stats.json \
  user_id=YOUR_USER_ID \
  submission_id=YOUR_SUBMISSION_ID

If UR5_NORM_STATS is unset, the client falls back to ${UR5_CHECKPOINT}/norm_stats.json.

Runtime Profile

From third_party/robochallenge_inference/configs/default.yaml → robot_profiles.ur5:

• action_type=leftpos, action_mode=relative, single_arm_target=eef
• ur5_anchor_pitch_zero=true, is_history=false
• speed=0.5, add_state=true
• attn: llm/action sdpa, vision flash_attention_2

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}
}