Modelpublic

ofcourseistillloveyou/act-so101-feed-me-vai-10ep-run1

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes7downloads

Model Card

ACT — SO-101 "feed me" — Vai, 10 episodes, run 1

Action Chunking Transformer (ACT) policy trained with LeRobot on `ofcourseistillloveyou/so-101-feed-me`.

—Operator: Vai
—Episodes: 10
—Run: 1

This snapshot is from step 2000 of an in-progress 20,000-step run. Subsequent checkpoints will be pushed as additional commits / tags (step-4000, step-6000, …, step-20000).

Training


Policy	`act` (52M params, ResNet18 vision backbone)
Dataset	10 episodes, 4,465 frames, 30 fps, 2 cameras (front + side), 6-DoF SO-101 follower
Batch size	16
Steps in this checkpoint	2,000
Mixed precision	AMP (bf16/fp16)
Hardware	NVIDIA RTX 3080 Laptop, 16 GB

Use on a real SO-101

bash

lerobot-record \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=my_follower \
  --robot.cameras="{ front: {type: opencv, index_or_path: /dev/video2, width: 640, height: 480, fps: 30}, side: {type: opencv, index_or_path: /dev/video4, width: 640, height: 480, fps: 30} }" \
  --dataset.repo_id=YOUR_USER/eval_so101_feed_me \
  --dataset.single_task="feed me" \
  --dataset.num_episodes=10 \
  --policy.path=ofcourseistillloveyou/act-so101-feed-me

Adjust index_or_path to whichever /dev/videoN matches your camera mounting.