xlerobot-coffee-md-real-force-loss

⚠️ PROVISIONAL — this naming grammar has not been ratified. It was generated on 2026-08-25 from a proposal whose open questions (docs/coffee/hub_taxonomy/NAMING_GRAMMAR.md §6) were never answered by a human. Leaf and repo names here may still change. Do not cite these paths in a paper, deck or script without checking that the grammar has since been signed off.

Coffee Automata — real domain. Part of the v2 taxonomy reorganization.

The legacy repos remain authoritative and untouched; this is a copy under a naming grammar that makes every experimental axis explicit.

Leaf grammar

<task>__<arch>__<N>cam__<platform>__<src><width>__<steps>[__<subset>]

Every field is always present, in this order, with no implicit defaults. The old scheme made _2cam optional, so a 1-camera model and a truncated name were indistinguishable.

field	values
arch	act act_v2 act_v3 smolvla pi0 dot — this also encodes the active-DoF loss version: v1=act, v2=act_v2, v3=act_v3, so it is never carried twice
platform	xlerobot-17dof / xlerobot-6dof-left / xlerobot-6dof-right / so101-6dof
src+width	none / est9 / est24 / gt12 (reserved — no model reads ground truth yet)
steps	training steps: 50k 75k 100k
subset	s567 (success 5+6+7) / redcup / eval / ds2 (second dataset build) / mixed

Two axes, not one. Force source (none/est/gt) is a property of the data; force use (none/in/loss) is a property of the model and names this repo. loss is a strict superset of in: it adds an auxiliary force-prediction term (hpi_pred_enabled).

Evaluation settings (n_action_steps 25 vs 100, checkpoint step) are not in these names — the same weights are evaluated at both, so they belong to the eval result.

PreTrainedPolicy.from_pretrained cannot read a subfolder; download the leaf and pass the dir:

from huggingface_hub import snapshot_download
from lerobot.policies.act.modeling_act import ACTPolicy
root = snapshot_download("IntelligentDecisionLab/xlerobot-coffee-md-real-force-loss", allow_patterns="g135_shared_t1_t3_t5__act__2cam__xlerobot-17dof__est24__100k/*")
policy = ACTPolicy.from_pretrained(f"{root}/g135_shared_t1_t3_t5__act__2cam__xlerobot-17dof__est24__100k")

Provenance — 5 leaves

new leaf	copied from
g135_shared_t1_t3_t5__act__2cam__xlerobot-17dof__est24__100k	xlerobot-coffee-model-real-d-force-closed-loop/g135_shared_t1_t3_t5_2cam_17dof
g35_shared_t3_t5__act__2cam__xlerobot-17dof__est24__100k	xlerobot-coffee-model-real-d-force-closed-loop/g35_shared_t3_t5_2cam_17dof
t1_place_cup__act__2cam__xlerobot-17dof__est24__100k	xlerobot-coffee-model-real-d-force-closed-loop/t1_place_cup_2cam_17dof
t3_cup_to_tray__act__2cam__xlerobot-17dof__est24__100k	xlerobot-coffee-model-real-d-force-closed-loop/t3_cup_to_tray_2cam_17dof
t5_tray_to_table__act__2cam__xlerobot-17dof__est24__100k	xlerobot-coffee-model-real-d-force-closed-loop/t5_tray_to_table_2cam_17dof