ฯ0.5 Pick3 baseline โ 3 trajectories, 80 epochs
EMA inference checkpoint at step 10,240, exactly 80 sampler epochs.
Global batch 8, one RTX A6000, 128 updates per sampler epoch.
The behavior subset contains 1031 H20 windows from training episodes:
pick3_20260826_211022_764 (ID 64), pick3_20260826_212741_569 (ID 78), pick3_20260827_161516_690 (ID 91).
Status-D epoch geometry includes its existing auxiliary Status mixture.
Current base/wrist RGB views, measured state and task text; no history memory.
Task: "Pick up the cube and place it on the plate exactly three times, lifting it clear between placements, then press the blue button."
The architecture is ฯ0.5 base with Cartesian8 LoRA, trained using the original
training-split normalization. Raw output shape is [20,8]: absolute
[x,y,z,qx,qy,qz,qw,gripper_open], positions in meters, unit XYZW quaternions,
and gripper 0=closed/1=open. Missing command labels stayed masked in training.
hf download fm-dev/pi05-pick3-baseline-overfit3-epoch80 --local-dir ./pick3-baseline-epoch80
The repository includes EMA params/, normalization and policy assets/,
matching model code/, requirements.txt, and load_model.py. Optimizer/resume
state is not included. Uniform32 and Status-D require real observed history;
Status-D additionally requires its held Writer inputs and causal context.
The packaged model was loaded on GPU and checked on nine recorded training
observations from all three episodes. See inference-check.json. These checks
verify loading and finite outputs; they do not establish real-robot success or
an online Writer rollout.