Datasets:

MarcelTorne
/

plate_toast_preferences

Real-robot trajectories for "put the toast in the plate" with human pairwise preference labels on multiple judgment axes. Built for reward-model / preference-learning research: every label is a comparison of two trajectories on one named axis, not a scalar score.

The trajectory data is a standard LeRobot v2.1 dataset, so it also loads directly as an imitation-learning dataset.


Episodes	271
Frames	83,209 (~1.5 h at 15 fps)
Preference pairs	1491
Preference labels (pair x axis)	2671
Judgment axes	43
Distinct instructions	1
Cameras	`agent_view` (third-person), `wrist` — both 224x224
Robot	Franka Panda, 7-DoF joint control + gripper

Episodes by kind

kind	episodes	frames
demo	148	31656
rollout	123	51553

demo episodes are human teleoperated demonstrations. rollout episodes are policy rollouts recorded in preference-collection sessions; those are the ones that vary in quality, which is what makes the comparisons informative.

Preference labels

preferences/pairs.parquet — one row per (pair, axis):

column	meaning
`pair_id`	unique id of the comparison
`episode_index_a`, `episode_index_b`	index into this LeRobot dataset
`episode_id_a`, `episode_id_b`	original source trajectory id
`axis`	the judgment axis being compared
`winner`	`A`, `B`, or `Equal`
`source`	`in_session` (the two rollouts of one session) or `cross` (arbitrary pair)
`source_dir`, `source_file`	provenance of the annotation
`annotator`	which annotator's label set this row comes from
`axis_set`	`fixed` (curated rubric reused across sessions) or `freeform` (annotator-invented axis names)
`overall_score_a/b`	1–4 Likert quality rating, where the annotator gave one (else null)
`succeeded_a/b`	task-success flag recorded at collection time, where available (else null)
`instruction`	instruction string logged with the comparison

preferences/pairs.jsonl is the same data with the axes nested per pair. preferences/episodes.parquet maps episode_index to the source trajectory.

Winner distribution: B 1175, A 1131, Equal 365.

Axes

axis	labels
`Overall quality`	1195
`hygiene`	140
`smoothness`	133
`speed`	120
`quality of plating`	119
`plating quality`	108
`fast`	101
`smooth`	90
`cleanliness`	78
`destroying the toast`	53
`toast destruction`	44
`destruction of the toast`	40
`placement on the plate`	36
`clean`	35
`toast demolishing`	34
`toast damage`	32
`stable`	27
`swift`	26
`overall quality`	23
`quickness`	22
`tidy`	21
`stability`	21
`toast deformation`	20
`environment destruction`	16
`damaging toast`	16
`hygienic`	16
`jerky movements`	13
`damaging the toast`	13
`deformation of the toast`	13
`quick`	11
`destroying to the toast`	8
`jerky motion`	8
`destruction of toast`	7
`quicker`	6
`fastness`	6
`dirtying the table`	5
`damage to the toast`	5
`destruction of environment`	3
`Overall Quality`	2
`destruction of the environment`	2
`placement on plate`	1
`lifting the toast`	1
`dropping spatula`	1

Pairs by source

source	directory	axis set	pairs
cross	`abhijnya/cross_preferences_burger`	freeform	700
cross	`am208/cross_preferences_burger`	freeform	730
in_session	`am208/preferences_burger`	freeform	61

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("MarcelTorne/plate_toast_preferences")
print(ds[0].keys())

# train a reward model on the pairwise labels
import pandas as pd
from huggingface_hub import hf_hub_download
prefs = pd.read_parquet(hf_hub_download("MarcelTorne/plate_toast_preferences", "preferences/pairs.parquet",
                                        repo_type="dataset"))
decisive = prefs[prefs.winner != "Equal"]           # drop ties
overall  = decisive[decisive.axis.str.lower().str.contains("overall")]

Both cameras are dtype: video. observation.state is [joint_0..joint_6, gripper], observation.ee is [x, y, z, roll, pitch, yaw], and action is the commanded [joint_position(7), gripper_position(1)].

Caveats

Frames are h264 (crf 16), not bit-exact. The source HDF5 held raw uint8 frames; re-encoding at the same 224x224 resolution shrinks the release ~100x at 41.8 dB PSNR (mean absolute error 1.1/255, 99th pct 7/255), measured on fold_pants cloth texture. Do not expect byte-level reproduction of the originals.
Equal is common on some axes. Annotators used it freely; filter deliberately.
freeform axis names are not a closed vocabulary. They were invented per session, so the same idea appears under several names (fast / speed, smooth / smoothness). Normalise before aggregating across sessions.
A pair can be labelled by more than one annotator. The two annotators' cross-pair sets are near-disjoint, but not perfectly; group by pair_id if you need unique pairs.
Instructions on rollout episodes sometimes carry reward conditioning (e.g. "put the toast in the plate, fast: 1.0, ...") because the policy that produced them was reward-conditioned. The LeRobot task string is the plain instruction; the logged string is kept in preferences/episodes.parquet.
No success labels on demo episodes — they are demonstrations, assumed good.