Benchmarking Robotic Manipulation
in the Real World
at Scale
with Any Robot at Any Time and Anywhere

"Manipulating atoms is more than manipulating bits."
Initiated from NIST Assembly Tasks to An Interconnected Network of Skills and Abilities for Real-world Manipulation.

KIT
THU
UT Austin
NVIDIA
UT Dallas
CU
KIT
CMU
DLR
KTH
Autodesk
MIT
Rice
NIST
UC Berkeley

A Global Infrastructure for Comparable Manipulation Research Supported By
U.S. National Institute of Standards and Technology

News

Latest updates from ManipulationNet — releases, results, and community highlights.

Task Released

"Language-conditioned Tabletop Manipulation" has been released

You can now register for this new task, "Language-conditioned Tabletop Manipulation", for benchmarking language-conditioned manipulation of everyday objects.

Check the task page →
Community News

ManipulationNet has received The First Submission

Team Linkerbot Research has completed all tasks on the NIST ATB M1 using a multi-fingered hand through teleoperation!

Check the Leaderboard →
Data

Example Scenes for the Grasping in Clutter Benchmark

We provide 100 example scenes for each difficulty level. Please use the local test client to evaluate your grasping performance.

Check the Github repo →
Task Released

"Grasping in Clutter" has been released

You can now register for this new task, "Grasping in Clutter", for benchmarking general robotic grasping.

Check the task page →
Policy Update

Submission policy has been updated

You can now submit more frequently for certain tasks. Please check the corresponding task page for more details.

Check the Github repo →
Client Update & New Task Added

The mnet-client has been updated

The mnet-client now supports the grasping_in_clutter task, which will be released next week. Please update the client before use.

Check the Github repo →
Client Update

The mnet-client has been updated

The new example tasks have been bridged to the mnet-client. Please update the client before use.

Check the Github repo →
Data

Example tasks for the Block Arrangement benchmark

We provide 100 example tasks for each task mode: language prompts, visual prompts, and visual–language prompts.

View example tasks →

Overview

ManipulationNet is a global, community-governed framework that hosts a wide range of real-world benchmarking tasks by which robots can be evaluated for their physical manipulation skills and embodied multimodal reasoning abilities.

It provides a persistent and scalable infrastructure to support comparable manipulation research across the globe and all active research questions of robot manipulation.

Distributed object sets for standardized task setups.

Physical benchmarks with reproducible setups across time and space

Your robot at your place. Your platform, at your convenience.

Maximize system, place, and time flexibility for skill-focused benchmarks

Distributed submissions with certifiable protocols.

Performance authenticity ensured for trustworthy and comparable results

Centralized evaluation by human judges.

Rapid, fair, and consistent evaluation and get you shine on the global leaderboards

How ManipulationNet Works?

Step 1
📋

Register for Tasks

Choose from available benchmark tasks and submit your registration request here.

Upon approval and as needed by specific task requirements, we will mail you the standardized object sets for constructing a reproducible task setup.

Step 2
🤖

Build Your Robotic System

Construct the robotic system to execute the benchmark tasks using the received object sets under the as-specified task protocol.

We encourage system-level innovation. Use any manipulator, any end-effector, any sensor, and any policy (model-based, data-driven, or even teleoperation) as you like.

Step 3
🎥

Submit Your Performance

After you have your system ready, you can submit your performance anytime through the mnet-client from your own place.

A ROS-compatible environment, stable internet connection, and an additional camera are required to record and submit your performance. No worries, you can always cancel a submission whenever needed. We only collect submissions you feel comfortable with.

Step 4
🏆

Shine before the World

Every performance submission will be centrally evaluated through standard task metrics and ranked on the global leaderboards.

Your performance video, execution logs, and any supplementary materials (if submitted, though not required) — such as the paper, technical report, and code — will be featured on the leaderboard.
Let the world see your achievement!

Tasks

ManipulationNet hosts a wide range of real-world benchmark tasks under the following tracks:
The Physical Skills Track evaluates physical interaction skills.
The Embodied Reasoning Track tests reasoning and multimodal grounding abilities.

Each task is designed to evaluate a specific skill or ability for real-world manipulation.

Open to collaboration!
If you have a benchmark task proposal that aligns with our mission, please contact us at [email protected]

Our Missions

ManipulationNet focuses on the most essential and fundamental, yet unsolved, skills and abilities for manipulation that make robots physically useful in the real world.

For Academia

If you are researchers who are pushing the frontier of robot manipulation research forward, ManipulationNet helps you to clearly identify the most up-to-date 1) real-world issues; 2) progresses; and 3) promising methodologies.

For Industry

If you are proud providers of robot solutions, whether you be a giant world-leader or a small and medium-sized business, ManipulationNet: 1) facilitates the exposure of your cutting-edge solutions; 2) connects you with the real-world needs; and 3) makes it possible for everyone to know what we can and cannot do for robotizing the world.

Meet the ManipulationNet Committee Members