pick-laptop-charger-act-ood

ACT policy trained on 59 of 83 episodes from the pick-laptop-charger-franka dataset, with the remaining 24 episodes held out as a spatial out-of- distribution (OOD) test set.

This model exists specifically to validate that the dataset generalizes to object positions the training set did not contain. It is not the model to use for deployment — for that, see pick-laptop-charger-act-v0 (trained on all 83 episodes).

Why this model exists

When evaluating an imitation-learning policy, a low replay-eval error on randomly-selected held-out episodes is a necessary but insufficient signal of dataset quality. It tells you the policy memorized the trajectories well — not that it generalizes to spatial conditions the demonstrator did not record.

To get a real signal, we group the dataset's 83 episodes by where the laptop charger was placed on the table (KMeans clustering on first-frame object bbox positions), pick the most spatially-isolated cluster as the OOD test set (24 episodes on the right side of the table), train ACT on the remaining 59, and replay-eval on the 24 held-out ones.

If the policy works on those 24 — episodes it never saw during training, all in a different region of the workspace — the dataset has captured a function of (image, state) → action, not memorized trajectories.

OOD eval results

Metric	In-distribution baseline (act-v0)	This model on 24 OOD	Ratio
pos_mae	2.8 mm	2.7 mm	0.96×
rot_mae	0.0088 rad	0.0094 rad	1.07×
gripper_mae	0.040	0.047	1.18×
gripper_timing_delta	n/a	-13 frames	early-close

Conclusion: the dataset generalizes. OOD positional performance is essentially indistinguishable from in-distribution performance on every metric we measured.

What this model does NOT prove

• Real-world deployment success: replay-eval is not closed-loop. The policy might still fail on a physical robot due to sim-to-real gap, controller latency, or lighting differences from the training video.
• Generalization beyond position: the OOD split is positional. The model was NOT validated on different lighting, camera angles, charger sizes, table colors, or backgrounds.
• Inter-operator variability: all 83 episodes were recorded by one human demonstrator. Cross-operator generalization is untested.

These are pre-deployment unknowns that need either sim eval, hardware tests, or more diverse data to resolve.

Training details

Same hyperparameters as act-v0:

Param	Value
Policy type	ACT (Action Chunking Transformer)
Training set	59 episodes (excluding spatial cluster 1)
Steps	20,000
Batch size	16
Optimizer	AdamW lr=1e-5
KL weight	10.0
Chunk size	100 frames
Vision backbone	ResNet18

Held-out training indices (output episode indices in the LeRobot dataset):

[24, 28, 29, 32, 33, 41, 42, 47, 48, 49, 50, 51, 58, 59, 60, 61,
 68, 69, 70, 75, 76, 79, 80, 82]

Same usage as act-v0 — see that model card for the load + patch + select action snippet.

When to use this model

• Reproducing the OOD validation in this dataset's README
• Comparing models trained on different splits
• Citation purposes when arguing about the dataset's generalization

For everything else, use pick-laptop-charger-act-v0 (the production model, trained on all 83 episodes).

Citation

Same source code and dataset as act-v0:

• Code: https://github.com/andresceballosm/humxai-glove-dataset
• Dataset: andresceballosm/pick-laptop-charger-franka

@misc{ceballos2026humxai_charger_ood,
  author = {Ceballos, Andrés Felipe},
  title  = {pick-laptop-charger-act-ood: positional OOD validation for the humxai laptop-charger skill dataset},
  year   = {2026},
  url    = {https://huggingface.co/andresceballosm/pick-laptop-charger-act-ood}
}