act-openarm-pick-and-place-ikea-red-block

ACT policy trained on the pick-and-place IKEA red block task (single right arm). 100k steps, chunk size 100, ResNet-18 vision backbone, absolute joint-space actions.

Trained on openarm-pick-and-place-ikea-red-block.

Summary

Field	Value
Policy	act
Robot	openarm
Task	pick-and-place
State / Action dims	8 / 8
Extra state inputs	—
Cameras	chest, waist, wrist_r (3)

Architecture

Field	Value
n_obs_steps	1
Chunk size	100
n_action_steps	100
Vision backbone	resnet18
Action normalization	MEAN_STD

Training

Field	Value
Source dataset	openarm-pick-and-place-ikea-red-block
Total episodes	250
Episodes used (train)	200
Splits	200 / 25 / 25 (train / val / test)
Training steps	100,000
Batch size	32
Optimizer	adamw, lr=1e-05
Image augmentation	no
W&B	placing-one-block-r4-fisheye / gf7cwi5s

Dataset

Field	Value
Source dataset	openarm-pick-and-place-ikea-red-block
Total episodes	250
Train / val / test	200 / 25 / 25 (train / val / test)
Trained with	anvil-embodied-ai

Task

Task description not available.

License

apache-2.0