Instructions to use Shiki42/s016-sortblocks-ctr-mask-act-step100000 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use Shiki42/s016-sortblocks-ctr-mask-act-step100000 with LeRobot:
- Notebooks
- Google Colab
- Kaggle
S016 Blocks Ranking โ ACT with IdleMask
This repository contains the inference checkpoint selected for Experiment E769-R001 at optimizer step 100,000. It is an inference-ready LeRobot ACT checkpoint with its configuration and preprocessing/postprocessing artifacts. Training optimizer state is not included.
Task and data
The policy was trained for the Blocks Ranking RGB-CTR simulation task using all 100 episodes and 68,438 frames from Shiki42/ctr-sortblocks-100ep-ctr, revision 9ee0f9d8e0700df5fac08454e7e213c4904d3024. The dataset was collected at 25 FPS with the centered_fovy90 camera preset and uniformly stratified relative start positions.
This is the CTR-with-IdleMask arm. Training consumed the dataset's prefix-only observation.arm_active_mask, aligned to each action window and combined with temporal padding.
Training
- Policy: official LeRobot ACT, version 0.4.4, source commit
8fff0fde7c79f23a93d845d1a50e985de01f8b8a. - CTR training source commit:
f9039a01ebb66dde3ed567608410bf170c34d6b1. - Fixed budget: 100,000 optimizer updates; effective batch size 8; seed 87431; gradient accumulation 1; chunk size 50.
- Inputs: RGB from top, left-wrist, and right-wrist cameras plus a 14-dimensional state; actions are 14-dimensional.
- Normalization: official ACT state/action mean and standard deviation, and ImageNet visual mean and standard deviation, computed and checked against valid rows of the pinned training revision.
- Checkpoint selection: the final step 100,000 checkpoint was selected before evaluation; no intermediate-metric selection or budget extension was used.
Verification and evaluation status
A fresh-process CPU reload passed for this checkpoint. All 234 state tensors matched exactly, and reset/action output was deterministic on a sample from the pinned dataset revision. This qualification used no GPU and performed no optimizer updates.
No held-out S016 evaluation result is claimed here. The shared held-out evaluation suite was not yet qualified when this card was prepared, so this card reports no success rate or collision result.
- Downloads last month
- 25