ฯ0.5 Shuffle baseline โ 3 trajectories, 100 epochs
Inference checkpoint at step 1,900, trained from ฯ0.5 base using the existing Cartesian8 LoRA recipe on three Shuffle training trajectories. Global batch is 8, one RTX A6000; the resumable sampler has 19 updates per epoch, so this checkpoint represents 100 sampler epochs and 15,200 sampled windows. The subset contains 158 overlapping H20 action windows.
| Original episode ID | Training trajectory |
|---|---|
| 111 | shuffle_20260908_170331_290 |
| 115 | shuffle_20260908_170704_618 |
| 131 | shuffle_20260908_171900_631 |
The model uses current base/wrist RGB images, measured Cartesian8 state, and the task instruction. It has no history memory or Writer. The instruction is:
After the cups are shuffled, press the button next to the cup hiding the cube.
Outputs are 20 absolute actions of shape [20, 8]:
[x, y, z, qx, qy, qz, qw, gripper_open], positions in meters and unit XYZW
quaternions. gripper_open uses 0=closed and 1=open. Action component 7 had no
known command supervision in this Shuffle subset; its raw predictions are
unsupervised. This is also recorded in assets/policy_metadata.json.
Download and load
hf download fm-dev/pi05-shuffle-baseline-overfit3-epoch100 --local-dir ./shuffle-epoch100
cd shuffle-epoch100
Use the included requirements.txt with a compatible Python/CUDA environment.
from load_model import load
policy = load()
params/ contains EMA inference parameters. assets/ contains normalization
and policy metadata, and code/ contains the corresponding model implementation.
Optimizer/resume state is not included in this inference export.
Validation
All 158 retained training observations produced finite [20, 8] actions after
reloading this checkpoint. These are training-set replay errors, not real-robot
success rates:
| Metric | Value |
|---|---|
| First-action position mean | 1.15 cm |
| First-action position P90 | 2.00 cm |
| First-action orientation mean | 1.06 degrees |
| Full H20 position mean | 0.89 cm |
See train-replay.json, training_config.json, and dataset_manifest.json for
the measured results and selected trajectories.