---
library_name: lerobot
base_model: lerobot-data-collection/folding_final
datasets:
- bohlt/openarm2-shirt-fold-phase-aligned-v1
tags:
- lerobot
- robotics
- pi0.5
- openarm2
- bimanual
- t-shirt-folding
- flow-matching
---

# Pi0.5 OpenArm2 T-shirt Folding — HF-Initialized 5K Fine-Tune

This is a **training candidate**, initialized from Hugging Face's final Pi0.5 folding policy and fine-tuned for 5,000 steps on phase-aligned Anvil OpenArm 2 T-shirt-folding demonstrations. It has not yet been performance-validated on hardware and is not deployment-approved.

## Provenance

| Item | Value |
|---|---|
| Base checkpoint | lerobot-data-collection/folding_final |
| Base revision | 695abe40dbf3aac04efda59c1501d748681fa0fb |
| Dataset | bohlt/openarm2-shirt-fold-phase-aligned-v1 |
| Dataset revision | 8411e3e85eaf3e482b4ccb1cac9d4fc02891305e |
| Training scope | Pi0.5 action expert only |
| Steps / batch size | 5,000 / 16 |
| Seed | 1000 |
| Action representation | Chunk-relative arm joints; grippers excluded from relative conversion |
| Checkpoint interval | 500 steps |
| Folding code revision | d7e2d09 |
| Anvil training revision | 8cb3b44 |
| W&B run | [v48pp7bv](https://wandb.ai/bensonlee55-none/openarm2-shirt-folding/runs/v48pp7bv) |

## Aggregate results

| Metric | Value |
|---|---:|
| Final logged train loss | 0.087000 |
| Final W&B train loss | 0.063375 |
| Final validation loss | 0.105837 |
| Final test loss | 0.084726 |
| Actuator-audit validation aggregate | 0.107207 |
| Actuator-audit test aggregate | 0.088134 |

![Aggregate loss curves](plots/loss_curves.png)

![Training diagnostics](plots/training_diagnostics.png)

## Per-actuator held-out loss

The grippers are the dominant residual-error channels. Their mean loss is **0.1757 validation / 0.1594 test**, compared with **0.0974 / 0.0779** across the 14 arm joints. The highest individual channel on both splits is **right_gripper.pos**. This is a normalized flow-matching loss, not an angle error or a measured joint-limit violation.

![Per-actuator losses](plots/per_actuator_losses.png)

| Actuator | Validation loss | Test loss |
|---|---:|---:|
| right_joint_1.pos | 0.103824 | 0.065850 |
| right_joint_2.pos | 0.104756 | 0.091587 |
| right_joint_3.pos | 0.093135 | 0.063144 |
| right_joint_4.pos | 0.098640 | 0.084986 |
| right_joint_5.pos | 0.068296 | 0.069760 |
| right_joint_6.pos | 0.069995 | 0.056355 |
| right_joint_7.pos | 0.109297 | 0.094386 |
| right_gripper.pos | 0.193641 | 0.183934 |
| left_joint_1.pos | 0.092994 | 0.075194 |
| left_joint_2.pos | 0.089928 | 0.063051 |
| left_joint_3.pos | 0.096795 | 0.070497 |
| left_joint_4.pos | 0.099758 | 0.089794 |
| left_joint_5.pos | 0.113030 | 0.094465 |
| left_joint_6.pos | 0.129621 | 0.084174 |
| left_joint_7.pos | 0.093949 | 0.088008 |
| left_gripper.pos | 0.157661 | 0.134964 |

Raw values and provenance are in [metrics](metrics/).

## Checkpoints

Numeric checkpoints are published every 500 steps from 000500 through 005000. The latest tag resolves to the complete step-5,000 checkpoint. Each checkpoint contains the policy, processors, embodiment contract, split information, and training state.

## Limitations and required validation

- Offline loss does not establish physical folding success.
- Prediction joint-limit violations must be audited against the repaired Anvil OpenArm 2 embodiment contract before hardware execution.
- Gripper channels have materially higher held-out loss and deserve focused rollout inspection.
- Trainer and actuator-audit losses use stochastic flow-matching samples and are not expected to be bit-identical.
- No hardware, deployment, or shadow-policy evaluation is claimed here.