daixianjie/controlled_anchor_v2_support_switch
Controlled Anchor — Task Suite V2 Task Suites 1, 2 and 3 together constitute Task Suite V2. Each suite is maintained in its own directory and preserves its own schema, scientific split and pairing metadata. The repository is not one merged LeRobot dataset. Task Suite 1: existing independently maintained data. Task Suite 2: controlled candy pouring; 1920 training + 132 held-out support/oracle episodes, 734887 frames at 20 Hz. All original JPEG images and float32 state/action… See the full description on the dataset page: https://huggingface.co/datasets/daixianjie/controlled_anchor_v2_support_switch.
Controlled Anchor — Task Suite V2
Task Suites 1, 2 and 3 together constitute Task Suite V2. Each suite is maintained in its own directory and preserves its own schema, scientific split and pairing metadata. The repository is not one merged LeRobot dataset.
- Task Suite 1: existing independently maintained data.
- Task Suite 2: controlled candy pouring; 1920 training + 132 held-out support/oracle episodes, 734887 frames at 20 Hz. All original JPEG images and float32 state/action values are preserved in Parquet.
- Task Suite 3: existing independently maintained data.
The viewer configuration above selects Task Suite 2 only. Its test split contains held-out expert demonstrations, not trained-policy evaluation rollouts. Do not use test frames for training or training normalization. Other suites' files remain separate.
from datasets import load_dataset
train = load_dataset("daixianjie/controlled_anchor_v2_support_switch",
name="task_suite_2", split="train")For LeRobot, download the selected suite directory and set its split directory as root; see Task Suite 2 loading and protocol. Support/query construction uses the frozen plan and split-qualified episode mappings, not the neutral per-trajectory task string alone.
