lucanunz/flowsteps

Paired A/B rollouts of six manipulation policies

4,800 teleoperation-free rollouts of six vision-language-action policies on a Franka Panda, across four manipulation tasks. Episodes are organised as 2,400 matched A/B pairs: the two arms of a pair run the same policy from the same initial scene with the same seed for noise sampling, differing only in the number of Euler (flow-matching) integration steps used to sample the action.

⚠️ Contains failures: read this before training on it

2,060 of the 4,800 episodes (43%) are failures. This is not a dataset of demonstrations.

For application like filtered Behavioural Cloning, filter on the success column:

from lerobot.datasets import LeRobotDataset, LeRobotDatasetMetadata

meta = LeRobotDatasetMetadata("lucanunz/flowsteps")
train = LeRobotDataset("lucanunz/flowsteps", episodes=meta.filter_episodes(lambda ep: ep["success"] == 1))

Episode metadata

Five columns beyond the LeRobot v3 defaults, in meta/episodes/*.parquet:

column	values	meaning
policy	dit, flower, lap, pi0.5, xr-0, xvla	which policy produced the rollout
num_euler_steps	1, 5, 10	integration steps used to sample each action — the A/B factor
success	0, 1	whether the task was accomplished, labelled by a human
pair_index	0–2399	the matched pair; identical on both arms
run_number	1, 2	The run order in the A/B pair (mostly metadata for sanity check, not meaningful for training)

Plus the v3 natives: episode_index, tasks, length, and per-episode stats/*.

Using the pairs

The two arms of pair k are always episode_index 2k and 2k+1, with the lower num_euler_steps first:

import pandas as pd
eps = meta.episodes.to_pandas()

# success rate per (policy, num_euler_steps)
print(eps.groupby(["policy", "num_euler_steps"])["success"].agg(["mean", "count"]))

# the discordant pairs — one arm succeeded, the other did not.
by_pair = eps.groupby("pair_index")["success"].agg(["min", "max"])
discordant = by_pair[by_pair["min"] != by_pair["max"]].index
print(f"{len(discordant)} discordant pairs of {len(by_pair)}")

run_number

Each pair's two arms ran back to back, and the execution order was randomised per pair. Included only as a sanity check for randomisation.

	ran first	ran second
num_euler_steps == 1	1242	1158
num_euler_steps == 5	401	399
num_euler_steps == 10	757	843

The A/B factor

Arm A is always 1 Euler step. Arm B is the policy's higher setting and the value is not the same for every policy:

policy	arm A	arm B	episodes
dit	1 step	10 steps	800
flower	1 step	5 steps	800
lap	1 step	10 steps	800
pi0.5	1 step	10 steps	800
xr-0	1 step	5 steps	800
xvla	1 step	10 steps	800

Tasks

instruction	episodes
Close the oven	1200
Open the drawer	1200
Open the oven	1200
Put the spoon in the drawer	1200

Dataset structure

episodes	4,800 (2,400 pairs × 2 arms)
frames	936,214
duration	21.7 h at 12 Hz
robot	Franka Panda + Robotiq 2F-85 gripper
control frequency	12 Hz
cameras	3 × 224×224 RGB: image, left_ego_image, wrist_image
format	LeRobotDataset v3.0

Features

• observation.state — float32[8]: [x, y, z, qx, qy, qz, qw, gripper]
• action — float32[8]: same layout, the commanded end-effector pose and gripper
• observation.images.{image,left_ego_image,wrist_image} — video, 224×224 RGB

Poses are the end-effector pose in the robot base frame.

The orientation is a unit quaternion in scalar-last order (qx, qy, qz, qw). Its sign is continuous within each episode (no q -> -q jumps between consecutive frames), with the first frame on the same hemisphere as the home pose (1, 0, 0, 0), i.e., a 180° rotation about the base x-axis (gripper pointing down).

The gripper channel is normalised and dimensionless, and the state and action channels are not on the same effective scale

	observed range (all 936,214 frames)	shape
observation.state[7]	[0.0001, 0.8340]	measured aperture; never exceeds 0.834
action[7]	[-0.1495, 1.0914]	commanded, unclipped; 76.1% near 0, 18.7% near 1, 5.2% between

The action is a binary open/close command that overshoots the nominal [0, 1] range, while the state is the aperture actually reached.

Camera intrinsics

In meta/robot_profile.json, which LeRobot v3 has no native slot for. The matrices describe the 224×224 images as stored, not the camera's native stream: frames were captured at 424×240 and resized to 224×224 by the policy observation pipeline, and the intrinsics were rescaled to match. That resize is anisotropic, so fx != fy (≈160 vs ≈282) — this is correct and encodes the non-square pixels, not a calibration error.

How the episodes were cut

Successful episodes end at the frame the task was completed. Rollouts were stopped by a human keypress and so originally ran on past the moment of success; those trailing frames were located by frame-level human review and removed. Failed episodes are unmodified.

Related

The same rollouts will also be published as part of the Oopsie Data community dataset of robotic manipulation failures. However, in that release the A/B pairing is lost.

Citation

If you use this data make sure to cite the related paper

@inproceedings{sujit2026when,
  title={When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies},
  author={Sujit, Shivakanth and Nunziante, Luca and Dossa, Rousslan Fernand Julien and Arulkumaran, Kai},
  booktitle={Conference on Robot Learning},
  year={2026},
  organization={PMLR}
}