Dataset readers¶

Generic LeRobot v3¶

The reader functions return Daft DataFrames directly:

from physical_ai_evals.datasets import (
    ALOHA,
    lerobot,
    lerobot_episodes,
    lerobot_tasks,
)

episodes = lerobot_episodes(ALOHA)
tasks = lerobot_tasks(ALOHA)
frames = lerobot(ALOHA, load_video_frames=False)

lerobot() decodes no video unless load_video_frames names a camera (or is True). lerobot_episodes() reads episode metadata only.

LeRobotSource binds a Hugging Face dataset to an exact revision:

from physical_ai_evals.datasets import LeRobotSource, lerobot_episodes

source = LeRobotSource("organization/dataset", "<40-character-commit>")
lerobot_episodes(source).show()

Daft 0.7.21 recursive Hugging Face globs lose the @revision component in listed paths. The wrapper therefore verifies that the repository head still equals the recorded commit before constructing the lazy read. It fails closed if the repository moved.

Recorded sources¶

Constant	Dataset	Revision
`ALOHA`	`lerobot/aloha_mobile_shrimp`	`6e828202059d2cc204b61ff968c232d202127a34`
`ABC_130K`	`lerobot/abc_130k_v3_train`	`68651e4929d9fb00f798937b2d62617cab5c771d`
`ABC_130K_SMOKE`	`lerobot/abc_130k_v3_smoke`	`b342a0ff262195d49bae3eece6e3f40c6e1dbe15`

LIBERO task plans¶

Each LIBERO constructor exposes its lazy executable episode plan directly:

import daft
from physical_ai_evals import libero_para, libero_pro

para = libero_para(task_ids=[3], episodes=1).rollouts
para.groupby("perturbation").agg(
    daft.col("bddl_path").count().alias("variants")
).show()

pro = libero_pro(
    "libero_spatial",
    perturbations=["lan"],
    episodes=1,
).rollouts
pro.select("task_key", "bddl_path", "init_path").show()

The constructors use Daft glob, regex, join, download, and expression nodes to turn the benchmark files into rows that the LIBERO runtime can execute.