anonymised-submission/XLERobot_Toy2Can_640x480

XLERobot Toy-to-Basket Mobile Manipulation

This dataset contains 101 teleoperated demonstrations of a mobile-manipulation task performed with a dual-arm XLERobot equipped with mecanum wheels. The robot approaches a table, picks up a plastic toy, and places it into a basket. The arm used in an episode is determined by the toy's initial side relative to the robot: a toy starting on the left is handled with the left arm, and a toy starting on the right is handled with the right arm. The other arm is intended to remain idle.

The dataset follows the LeRobot v3.0 format and contains synchronized robot state, action, and RGB video observations.

Dataset summary

Property	Value
Robot	Dual-arm XLERobot, mecanum-wheel version
Episodes	101
Frames	47,357
Recording frequency	15 Hz
Recorded duration	approximately 52 min 37 s
Tasks	1
RGB cameras	3
Image resolution	640 x 480
Observation state	17 dimensions
Action	15 dimensions
LeRobot format	v3.0

The stored task instruction is:

Pick up the toy and place it into the can.

Here, can refers to the receiving basket used during collection.

Collection protocol

At the beginning of an episode, the target toy is visible in the central camera and is generally within approximately 0.5 m of the robot. The initial relation between the robot and the toy is lightly randomized. A demonstration normally consists of:

1. approaching the table and target with the mecanum base;
2. stopping the base near the target;
3. reaching and grasping with the arm on the toy's starting side;
4. transporting the toy to the basket; and
5. releasing the toy into the basket.

The two arms were teleoperated using two SO-101 leader arms. The mobile base was controlled from a keyboard. Collection used nominal base speed limits of 0.1 m/s for translation and 10 degrees/s for rotation.

Episodes 0-49 and 50-100 were collected in two collection sessions by different operators. This provides some operator variation, but collection order, operator, starting configurations, and hardware state are confounded and should not be treated as independently controlled factors.

Observations and actions

RGB observations

Feature	Description
observation.images.main	Central Intel RealSense RGB view
observation.images.left_wrist	Camera attached to the left arm
observation.images.right_wrist	Camera attached to the right arm

All distributed videos are 640 x 480 at 15 fps, encoded as AV1 with yuv420p pixel format and no audio.

The central camera stream was retained at its recorded 640 x 480 resolution. The original wrist streams were 848 x 480. To produce a common tensor shape without cropping their horizontal field of view, wrist frames were resized to 640 x 362 with aspect ratio preserved and vertically letterboxed to 640 x 480. Image statistics were recomputed after conversion. Frame counts, frame rates, and output dimensions were validated for every video file.

State

observation.state is a 17-dimensional vector containing:

• six joint positions for the left arm, including its gripper;
• six joint positions for the right arm, including its gripper;
• two observed head joint positions; and
• three mobile-base velocities: x.vel, y.vel, and theta.vel.

Action

action is a 15-dimensional vector containing:

• six position commands for the left arm, including its gripper;
• six position commands for the right arm, including its gripper; and
• three mobile-base velocity commands: x.vel, y.vel, and theta.vel.

The head joints are observed but are not commanded by the policy interface.

The complete feature names, shapes, codecs, and storage templates are available in meta/info.json.

Active-arm convention

The active arm is not stored as a separate observation or episode field. It can be inferred reliably from the two six-joint arm trajectories. A post-collection motion-range check identified 51 left-arm and 50 right-arm episodes; all 101 episodes showed clear separation between the active and idle arm under that check.

Users studying multimodal policies may wish to derive an active_arm label or otherwise condition the policy on the intended arm. Without such conditioning, visually similar initial states can correspond to different valid action modes.

Loading with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    "anonymised-submission/XLERobot_Toy2Can_640x480"
)

print(dataset.num_episodes)  # 101
print(dataset.num_frames)    # 47357

Intended uses

The dataset is intended for research on:

• visuomotor imitation learning;
• action chunking and multimodal action prediction;
• mobile manipulation with coupled base and arm actions;
• dual-arm mode selection; and
• transfer to expanded robot state and action spaces.

Limitations and known data characteristics

• The dataset covers one task, one primary plastic toy, one robot, and one collection environment. It should not be interpreted as evidence of general object-manipulation capability.
• All 101 episodes belong to the train split. The dataset does not define a held-out evaluation split.
• There are no explicit success, task-stage, object-pose, or active-arm labels.
• Six episodes have unusually long leading idle periods: episodes 76, 82, 83, 90, 97, and 99 contain approximately 6.5-14.7 seconds before meaningful motion. These frames have intentionally not been removed from the released version.
• The two collection blocks have different starting-state distributions. The largest observed shift is in initial gripper targets, with additional variation in wrist and shoulder starting values.
• The active-arm convention is approximately balanced overall, but object position, selected arm, operator, and episode order were not independently randomized as a factorial experiment.
• The wrist views contain black letterbox regions introduced during resolution standardization. No spatial crop was applied.
• The dataset contains no audio or operator identity metadata. RGB streams can nevertheless contain incidental details of the physical collection environment.

These properties should be considered when constructing train/validation splits and when interpreting generalization results. In particular, a random frame-level split would leak temporally adjacent samples; evaluation should be performed using complete held-out episodes or newly collected physical trials.

License

The dataset is released under the Apache License 2.0.

Citation

The associated work is under anonymous review. A citation record will be added after the review process.