This dataset provides rich multi-modal demonstrations in LeRobot format, covering 20 real-robot tasks and 3 simulation tasks. Beyond standard end-effector trajectories, we provide joint positions, velocities, currents, camera views, and mobile-base states — giving participants the freedom to explore diverse input representations.

Changelog

2026-05-12

Added

Three-level language annotations for the 15 tabletop real-robot tasks (execution_reasoning x10 + semantic_reasoning x5) under language_annotations/real/<category>/<task>.jsonl. Each episode now has a short task name, a detailed action description, and a scene + action description (see the Language Annotations section).

Changed

Replaced the earlier 60 / 50 / 50-episode simulation data with high-quality refreshed demonstrations for the three sim tasks. The updated simulation set has larger episode counts (660 / 501 / 501) recorded with an ex001_6r dual-arm robot, and the state/action layout is now a combined 28D = end-effector (14D) + joint (14D). Front-camera resolution is now 720×1280 (wrist cameras remain 480×640). press_button_in_order now has 11 task variants (different button-color orderings).

Unchanged

All real/<category>/<task>/{meta,data,videos} trajectory folders are untouched.

2026-03-17

Initial release

20 real-robot tasks (10 execution_reasoning + 5 semantic_reasoning + 5 mobile_manipulation) in LeRobot v2.1 format.
3 simulation tasks (60 / 50 / 50 episodes).
3 synchronized camera streams (front + left wrist + right wrist), 20 Hz recording.

Dataset Structure

maniparena-dataset/
├── real/
│   ├── execution_reasoning/        (10 tasks, ~5,000 episodes)
│   ├── semantic_reasoning/         (5 tasks, ~2,800 episodes)
│   └── mobile_manipulation/        (5 tasks, ~2,900 episodes)
└── sim/
    ├── press_button_in_order/      (660 episodes, 11 task variants)
    ├── put_blocks_to_color/        (501 episodes)
    └── pick_fruits_into_basket/    (501 episodes)

Each task folder follows LeRobot format:

<task>/
    meta/info.json
    meta/tasks.jsonl
    data/chunk-000/episode_000000.parquet
    videos/chunk-000/
        observation.images.faceImg/episode_000000.mp4
        observation.images.leftImg/episode_000000.mp4
        observation.images.rightImg/episode_000000.mp4

Real Robot Data

Tabletop tasks (Execution Reasoning + Semantic Reasoning) have 56-dimensional state/action vectors. Mobile Manipulation tasks have 62-dimensional state/action vectors (56D + 6D mobile extras).

Dimension Layout

End-effector (index 0–13, 14D):

Index	Key	Dim	Description
0–2	`follow_left_ee_cartesian_pos`	3	Left arm position (x, y, z)
3–5	`follow_left_ee_rotation`	3	Left arm rotation (roll, pitch, yaw)
6	`follow_left_gripper`	1	Left gripper open/close
7–9	`follow_right_ee_cartesian_pos`	3	Right arm position (x, y, z)
10–12	`follow_right_ee_rotation`	3	Right arm rotation (roll, pitch, yaw)
13	`follow_right_gripper`	1	Right gripper open/close

Coordinate system: +x forward, +y left, +z up.

Joint — left arm (index 14–34, 21D):

Index	Key	Dim	Description
14–20	`follow_left_arm_joint_pos`	7	Left arm joint positions (6 joints + gripper)
21–27	`follow_left_arm_joint_dev`	7	Left arm joint velocities (6 joints + gripper)
28–34	`follow_left_arm_joint_cur`	7	Left arm joint currents (6 joints + gripper)

Joint — right arm (index 35–55, 21D):

Index	Key	Dim	Description
35–41	`follow_right_arm_joint_pos`	7	Right arm joint positions (6 joints + gripper)
42–48	`follow_right_arm_joint_dev`	7	Right arm joint velocities (6 joints + gripper)
49–55	`follow_right_arm_joint_cur`	7	Right arm joint currents (6 joints + gripper)

The last element (index 20, 27, 34, 41, 48, 55) in each 7D joint group is the gripper value.

Mobile manipulation extras (index 56–61, mobile tasks only, 6D):

Index	Key	Dim	Description
56–57	`head_actions`	2	Head rotation (yaw, pitch)
58	`height`	1	Lift mechanism height
59–61	`velocity_decomposed_odom`	3	Chassis velocity (vx, vy, angular velocity)

Tabletop tasks = 56D (index 0–55). Mobile Manipulation tasks = 62D (index 0–61).

Task List — Real Robot

Execution Reasoning (10 tasks):

Task	Episodes	Key Challenge
`arrange_cup_inverted_triangle`	528	Multi-object spatial planning
`put_spoon_to_bowl`	525	Precision grasping, varied shapes
`put_glasses_on_woodshelf`	513	Fragile object handling
`put_ring_onto_rod`	517	Sub-cm insertion precision
`put_items_into_drawer`	510	Multi-object coordination
`pick_items_into_basket`	532	Adaptive grasping
`pour_water_from_bottle`	526	Force control, liquid dynamics
`insert_wireline`	530	Contact-rich, mm-level accuracy
`put_stationery_in_case`	525	Multi-object organization
`put_blocks_to_color`	451	Color-zone matching

Semantic Reasoning (5 tasks):

Task	Episodes	Key Challenge
`sort_headphone`	515	Recognize headphone type
`classify_items_as_shape`	545	Map objects to shape categories
`press_button_in_order`	538	Color-button mapping + sequence
`pair_up_items`	540	Match pairs by pattern
`pick_fruits_into_basket`	645	Fruit vs. non-fruit distinction

Mobile Manipulation (5 tasks):

Task	Episodes	Key Challenge
`put_clothes_in_hamper`	540	Navigate + pick clothes
`hang_up_picture`	576	Navigate to wall + hang
`organize_shoes`	595	Navigate + arrange on rack
`put_bottle_on_woodshelf`	630	Navigate to shelf + place
`take_and_set_tableware`	531	Navigate + set table

Simulation Data (28D)

Simulation demonstrations contain 28-dimensional observation.state and action vectors, combining end-effector (14D) and joint (14D) representations of the same trajectories produced by an ex001_6r dual-arm robot.

Dimension Layout

End-effector (index 0–13):

Index	Key	Dim	Description
0–2	`left_pos_x/y/z`	3	Left arm EE position (x, y, z)
3–5	`left_rot_x/y/z`	3	Left arm EE rotation (roll, pitch, yaw)
6	`left_gripper`	1	Left gripper
7–9	`right_pos_x/y/z`	3	Right arm EE position (x, y, z)
10–12	`right_rot_x/y/z`	3	Right arm EE rotation (roll, pitch, yaw)
13	`right_gripper`	1	Right gripper

Joint (index 14–27):

Index	Key	Dim	Description
14–19	`left_arm_joint1..6`	6	Left arm joint positions
20	`left_arm_gripper`	1	Left gripper (joint side)
21–26	`right_arm_joint1..6`	6	Right arm joint positions
27	`right_arm_gripper`	1	Right gripper (joint side)

The first 14 dimensions (EE) are directly compatible with real-robot index 0–13.

Task List — Simulation

Task	Episodes	Task variants	Real-robot Counterpart
`press_button_in_order`	660	11 color orders	`semantic_reasoning/press_button_in_order`
`pick_fruits_into_basket`	501	1	`semantic_reasoning/pick_fruits_into_basket`
`put_blocks_to_color`	501	1	`execution_reasoning/put_blocks_to_color`

press_button_in_order has 11 task variants (different button-color orderings) indexed via task_index in each parquet; see meta/tasks.jsonl.

Camera Views

All tasks include 3 synchronized camera streams. Real-robot videos are 480×640. Updated simulation videos use 720×1280 for the front camera and 480×640 for wrist cameras.

Camera	Key	Description
Front	`observation.images.faceImg`	Third-person overhead view
Left wrist	`observation.images.leftImg`	Left arm wrist-mounted camera
Right wrist	`observation.images.rightImg`	Right arm wrist-mounted camera

Recording Frequency

All data is recorded at 20 Hz.

Language Annotations

For the 15 tabletop tasks (10 execution_reasoning + 5 semantic_reasoning), we provide three-level language annotations in a dedicated top-level folder:

language_annotations/
└── real/
    ├── execution_reasoning/
    │   ├── arrange_cup_inverted_triangle.jsonl
    │   ├── insert_wireline.jsonl
    │   └── ... (10 tasks)
    └── semantic_reasoning/
        ├── classify_items_as_shape.jsonl
        └── ... (5 tasks)

Each .jsonl file follows the LeRobot episodes.jsonl schema (one JSON object per episode, aligned with real/<category>/<task>/data/... via episode_index). The tasks field contains three strings:

Index	Type	Description
`tasks[0]`	Short task	One-sentence task name (e.g. "Arrange the cups into an inverted triangle.") — identical to `real/<task>/meta/episodes.jsonl`
`tasks[1]`	Detailed action	Step-by-step description of what the robot does, including object attributes and target configuration
`tasks[2]`	Scene + action	Full scene description (objects, layout, background) followed by the detailed action

mobile_manipulation and sim tasks are not covered yet — their real/<task>/meta/episodes.jsonl and sim/<task>/meta/episodes.jsonl still contain only the short task name.

import json, pandas as pd

# Load language annotations
with open("language_annotations/real/execution_reasoning/arrange_cup_inverted_triangle.jsonl") as f:
    annos = {json.loads(line)["episode_index"]: json.loads(line)["tasks"] for line in f}

# Join with trajectory data
df = pd.read_parquet("real/execution_reasoning/arrange_cup_inverted_triangle/data/chunk-000/episode_000000.parquet")
short, action, scene_and_action = annos[0]

Quick Usage

import pandas as pd
import numpy as np

# Load one episode
df = pd.read_parquet("real/execution_reasoning/put_blocks_to_color/data/chunk-000/episode_000000.parquet")

state = np.stack(df["observation.state"].tolist())  # (T, 56) for tabletop real, (T, 62) for mobile, (T, 28) for sim
action = np.stack(df["action"].tolist())

# EE data (first 14 dims — same layout for real and sim)
ee_left = state[:, 0:7]    # xyz(3) + rpy(3) + gripper(1)
ee_right = state[:, 7:14]

# Joint data is available for real-robot trajectories.
left_arm_joint_pos = state[:, 14:21]   # 6 joints + gripper
right_arm_joint_pos = state[:, 35:42]  # 6 joints + gripper

Citation

If you find this dataset useful in your research, please consider citing:

@misc{sun2026maniparena,
      title={ManipArena: Comprehensive Real-world Evaluation of Reasoning-Oriented Generalist Robot Manipulation},
      author={Yu Sun and Meng Cao and Ping Yang and Rongtao Xu and Yunxiao Yan and Runze Xu and Liang Ma and Roy Gan and Andy Zhai and Qingxuan Chen and Zunnan Xu and Hao Wang and Jincheng Yu and Lucy Liang and Qian Wang and Ivan Laptev and Ian D Reid and Xiaodan Liang},
      year={2026},
      eprint={2603.28545},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.28545},
}

License

Apache License 2.0

Downloads last month: 5,912

Total file size:

253 GB

Paper for ManipArena/maniparena-dataset

ManipArena: Comprehensive Real-world Evaluation of Reasoning-Oriented Generalist Robot Manipulation

Paper • 2603.28545 • Published Mar 30 • 1