HiveBoard Website
This repository hosts the official project website for HiveBoard: An Open, Modular, 3D-Printed Benchmark of Industrial Mechanisms for Robotic and Prosthetic Manipulation.
Live Website: hiveboard-bench.github.io
Main Project Repository: github.com/hiveboard-bench/HiveBoard
Website Purpose
The purpose of this website is to serve as the interactive companion and presentation portal for the HiveBoard benchmark:
- Interactive 3D & Physics Viewer: Allows visitors to view and interactively simulate the attachments directly in the browser using Three.js and WebAssembly-powered MuJoCo.
- Attachment & Task Catalog: Details 12 functional 3D-printed attachment designs evaluated in 13 conditions across three manipulation skill categories: Torque, Precision, and Composed assembly.
- Simulation Compatibility: Provides information and direct access to digital assets for simulators including MuJoCo, Isaac Sim (USD), and standard URDFs.
- Experimental Results & Media: Showcases evaluation results and video demonstrations across diverse robotic embodiments (fixed-base arms, quadruped manipulators, VR-teleoperated robots, and wearable prosthetic hands).
- Documentation & Evaluation Protocol: Includes the full HiveBoard documentation, contribution guides, and browser-based 65-trial evaluation runner at
/hivedocs/.
Tech Stack
Getting Started
Prerequisites
Make sure you have Node.js (version 18+ recommended) installed.
Installation
Clone the repository and install the dependencies:
git clone https://github.com/hiveboard-bench/hiveboard-bench.github.io.git
cd hiveboard-bench.github.io
npm install
Local Development
Start the local Vite development server:
npm run dev
Open your browser at http://localhost:5173 (or the URL printed in your terminal).
Build for Production
To create an optimized production build in the dist/ directory:
npm run build
You can preview the built site locally with:
npm run preview
Deployment
Deploy to GitHub Pages:
npm run deploy
Modifying & Authoring Trajectories
The simulation viewer replays deterministic task-space trajectories (approach, grasp, actuate, release, retreat) solved with inverse kinematics and verified against MuJoCo physics. You can modify these trajectories in two ways:
1. Interactive Browser-Based Editor
The repository includes a live visual editor to adjust waypoints directly in 3D:
- Start the editor backend server:
python3 tools/traj_edit.py
- Start the Vite dev server (in a separate terminal):
npm run dev
- Open the editor in your browser:
Navigate to http://localhost:5173/sim/hiveboard-sim.html?edit=1
- Interact & Save:
- Click and drag the keyframe beads along the motion path (or select a bead and use arrow keys to nudge coordinates).
- The backend re-solves IK and simulates acceptance replay in real-time.
- Click Save to persist waypoint offsets to
tools/traj_edits.json and update the robot's .traj.json file.
2. Programmatic Authoring via Python
(This compiles MuJoCo scenes, solves IK, and runs acceptance tests across all attachments).
NVIDIA cuMotion for FR3 and Spot
The FR3 and Spot trajectory builders use NVIDIA's standalone cuMotion package.
It requires a supported NVIDIA GPU and a matching CUDA release. On Ubuntu 24.04
with CUDA 13, install the project-local Python environment and official NVIDIA
wheel with:
tools/install-cumotion.sh
.venv/bin/python tools/test_cumotion_integration.py
.venv/bin/python tools/test_cumotion_tasks.py
The installer defaults to cuMotion 1.1.0 for CUDA 13.0. For Ubuntu 22.04 with
CUDA 12.6, run CUMOTION_CUDA_VERSION=12.6 tools/install-cumotion.sh. NVIDIA's
binary package supports Python 3.10–3.14; the installer picks the matching
wheel for python3. See NVIDIA's system requirements and installation
guide
before choosing a package for another platform.
To run the planner directly, pass the robot's target joint positions. The
planner defaults the starting state to that robot's cuMotion home pose:
.venv/bin/python tools/cumotion_planner.py --robot fr3 \
--goal -0.65 -0.04 0.64 -2.08 0.02 2.05 0.77
To regenerate the simulator scenes and rerun task acceptance for FR3 or Spot,
provide the HiveBoard asset checkout:
.venv/bin/python tools/build-sim-assets.py \
--hiveboard /path/to/HiveBoard/Simulation --robot fr3
.venv/bin/python tools/build-sim-assets.py \
--hiveboard /path/to/HiveBoard/Simulation --robot spot
The cuMotion Python runtime is a GPU binary and is not installed by the web
site's Node.js dependencies. The browser continues to play the generated
MuJoCo trajectories; Python cuMotion is needed only to regenerate them.
ANYmal-D + DynaArm + Robotiq 2F-140
The anymal model is available in public/sim/models/anymal.xml, with editable
starter trajectories for the ball valve, light bulb, and circuit breaker. To start editing:
pip install mujoco numpy
python3 tools/traj_edit.py --robot anymal
# In another terminal: npm run dev
Open http://localhost:5173/sim/hiveboard-sim.html?edit=1&robot=anymal.
The initial paths are drafts: their contact replays do not pass task acceptance
yet. They are starting points for waypoint/joint editing, not successful demos.
To regenerate only this robot from the Isaac assets (run from this directory):
pip install mujoco numpy usd-core fast-simplification
python3 tools/build-sim-assets.py --robot anymal \
--isaaclab-repo ../.. --hiveboard ../HiveBoard/Simulation
python3 tools/check-anymal.py --isaaclab-repo ../..
--isaaclab-repo points to the isaaclab-hiveboard checkout. NVIDIA USD layers
are downloaded into .cache/anymal-usd on the first build; --usd-cache can
select an existing download. Viewing and editing the generated model needs no
USD installation or NVIDIA connection. Full asset builds also include ANYmal.
The local assets/anymal/usd/dynaarm.usd and its configuration/ layers must
exist before rebuilding. These are ignored generated files in the Isaac repo:
spawning ANYmal there once creates them from the committed DynaArm URDF and
the duatic_dynaarm submodule. Regenerate them after changing that URDF.
The conversion uses the ANYmal-D USD, the local corydoras12 DynaArm USD/URDF,
and the Isaac Robotiq 2F-140 USD. It preserves the six arm joint names and limits,
the mount at (0, 0, 0.12) with 180° yaw, and the identity flange-to-palm weld.
The TCP is 0.20 m along the palm's +Z, with the website's +Z approach convention
(Isaac's canonical TCP uses a different orientation offset). Gripper commands
are radians: 0 open, 0.7 closed, with the inner fingers following the drive at -1.
The base and legs are fixed in the Isaac standing pose, as required by the
existing arm trajectory editor. MuJoCo uses position servos, gravity compensation,
convex mesh contacts and explicit four-bar loop/mimic constraints. Robot self
collisions are disabled, matching the Isaac assembly. Visual meshes are simplified
and texture maps replaced with solid colors. This is a manipulation-authoring
model, not a locomotion model or a dynamics-equivalent PhysX conversion.
See the converter for asset URLs and
the validation script for URDF forward-kinematics,
gripper linkage, editor and packaging checks. The generated model was also
checked with the bundled MuJoCo WASM 3.11.0 runtime.
Repository Structure
├── assets-src/ # Source assets and 3D models
├── docs/ # VitePress documentation and evaluation runner
├── public/ # Static assets served directly (textures, models, videos)
├── src/ # Frontend source scripts and utilities
├── tools/ # Python helper scripts for model compression and asset optimization
├── index.html # Main HTML page and presentation layout
├── viewer.js # MuJoCo WebAssembly and Three.js 3D interactive viewer logic
├── style.css # Website styling and responsive design rules
└── package.json # Node.js dependencies and scripts
Open Calls for Contributions
We invite laboratories, individual researchers, and teams to contribute manipulation datasets or develop functional attachments for benchmark extensions:
- Learning Datasets: Robot demonstrations and rollouts with synchronized states, actions, camera observations, and task outcomes. Use DataHive or agreed custom formats.
- Activities of Daily Living (ADL): Functional mechanisms representing everyday manipulation tasks with repeatable initial states and observable success conditions.
- Bimanual Manipulation: Attachments requiring coordinated manipulation by two hands or end-effectors.
Substantial accepted datasets and functional attachment packages can establish eligibility for co-authorship on future papers. See the submission requirements and credit policy. Contact Ricardo V. Godoy before starting design or collection.
Citation
If you find HiveBoard useful in your research, please cite:
@article{godoy2024hiveboard,
title = {HiveBoard: An Open, Modular, 3D-Printed Benchmark of Industrial Mechanisms for Robotic and Prosthetic Manipulation},
author = {Godoy, Ricardo V. and
de Souza, Enzo F. and
Angarola, Matheus P. and
Negri, Juliano and
Aléssio, João H. and
van Halst, Victor and
Milazzo, Giuseppe and
Capezzuto, Gianluca and
Tommaselli, Felipe and
Affonso, Francisco and
Marsicano, Jo{\~a}o A. and
de Lange, Rudy De-Xin and
Vijayan, Aravind Elanjimattathil and
Ramírez Sánchez, Amy M. and
Baptista, Rafael R. and
van Berge, Meiko A. and
Chowdhary, Girish and
Bezerra, Ranulfo and
Lahr, Gustavo J. G. and
Gerez, Lucas Ferrari and
Bicchi, Antonio and
Becker, Marcelo},
journal = {arXiv preprint},
year = {2024}
}
