KinDER

KinDER: A physical reasoning benchmark for robot learning and planning.
Robotics: Science and Systems (RSS), 2026

Website

See https://prpl-group.com/kinder-site/ for documentation and tutorials.

Requirements

1. Python >=3.10, <3.13
2. Tested on MacOS 13-15, Ubuntu 20.04, Ubuntu 22.04, Ubuntu 24.04, and Windows 10 (but we aim to support most platforms)

Installation

From PyPI

pip install kindergarden   # all environments (PyBullet, MuJoCo, pygame, ...)

Installing a single backend (no PyBullet/MuJoCo)

The default install pulls every backend. To install just one — for example the kinematic2d environments, which need neither PyBullet nor MuJoCo — install the package without its dependencies, then the backend's requirements file:

pip install --no-deps kindergarden
pip install -r https://raw.githubusercontent.com/Princeton-Robot-Planning-and-Learning/kindergarden/main/requirements/kinematic2d.txt

Requirements files are provided for each backend: kinematic2d, dynamic2d, kinematic3d, kinematic3d_v2, dynamic3d (under requirements/). The two-step form is needed because pip extras can only add dependencies, never remove the backends already pulled in by the base install.

The prpl_kinematics backend

The kinematic3d_v2 environments — currently VegaMotion3D, which uses the Dexmate Vega — are built on prpl_kinematics rather than pybullet_helpers. Because those two packages overlap heavily and are both large, a base install pulls in neither pair; add prpl_kinematics on top:

pip install "kindergarden[prpl-kinematics]"

Environments whose backend is missing are silently skipped at registration, so kinder.make("kinder/VegaMotion3D-v0") raises a gymnasium lookup error until prpl_kinematics is installed. To install the backend alone, use requirements/kinematic3d_v2.txt with the two-step form above.

From Source

We strongly recommend uv, but other standard setups work too.

git clone https://github.com/Princeton-Robot-Planning-and-Learning/kindergarden.git
cd kindergarden
uv pip install -e ".[develop]"   # all dependencies + dev tools

Or install only one backend (no PyBullet/MuJoCo):

uv pip install --no-deps -e .
uv pip install -r requirements/kinematic2d.txt

To check the installation, run ./run_ci_checks.sh. It should complete with all green successes.

Usage Example

Basic Usage (Gym API)

import kinder
kinder.register_all_environments()
env = kinder.make("kinder/Obstruction2D-o3-v0")  # 3 obstructions
obs, info = env.reset()  # procedural generation
action = env.action_space.sample()
next_obs, reward, terminated, truncated, info = env.step(action)
img = env.render()  

Object-Centric States

All environments in KinDER use object-centric states. For example:

from kinder.envs.kinematic2d.obstruction2d import ObjectCentricObstruction2DEnv
env = ObjectCentricObstruction2DEnv(num_obstructions=3)
obs, _ = env.reset(seed=123)
print(obs.pretty_str())

Here, obs is an ObjectCentricState, and the printout is:

############################################################### STATE ###############################################################
type: crv_robot           x         y    theta    base_radius    arm_joint    arm_length    vacuum    gripper_height    gripper_width
-----------------  --------  --------  -------  -------------  -----------  ------------  --------  ----------------  ---------------
robot              0.885039  0.803795  -1.5708            0.1          0.1           0.2         0              0.07             0.01

type: rectangle           x         y    theta    static    color_r    color_g    color_b    z_order      width     height
-----------------  --------  --------  -------  --------  ---------  ---------  ---------  ---------  ---------  ---------
obstruction0       0.422462  0.100001        0         0       0.75        0.1        0.1        100  0.132224   0.0766399
obstruction1       0.804663  0.100001        0         0       0.75        0.1        0.1        100  0.0805652  0.0955062
obstruction2       0.559246  0.100001        0         0       0.75        0.1        0.1        100  0.12608    0.180172

type: target_block          x         y    theta    static    color_r    color_g    color_b    z_order     width    height
--------------------  -------  --------  -------  --------  ---------  ---------  ---------  ---------  --------  --------
target_block          1.20082  0.100001        0         0   0.501961          0   0.501961        100  0.138302  0.155183

type: target_surface           x    y    theta    static    color_r    color_g    color_b    z_order     width    height
----------------------  --------  ---  -------  --------  ---------  ---------  ---------  ---------  --------  --------
target_surface          0.499675    0        0         1   0.501961          0   0.501961        101  0.180286       0.1
#####################################################################################################################################

For compatibility with baselines, the observations provided by the main environments are vectors. It is easy to convert between vectors and object-centric states. For example:

import kinder
kinder.register_all_environments()
env = kinder.make("kinder/Obstruction2D-o3-v0")
vec_obs, _ = env.reset(seed=123)
object_centric_obs = env.observation_space.devectorize(vec_obs)
recovered_vec_obs = env.observation_space.vectorize(object_centric_obs)

Noisy Observation and Action Wrappers

KinDER provides Gymnasium-compatible wrappers for adding stochasticity to observations and actions:

import kinder
kinder.register_all_environments()
env = kinder.make("kinder/Obstruction2D-o3-v0")
env = kinder.NoisyObservation(env, noise_std=0.05)  # Gaussian noise on observations
env = kinder.NoisyAction(env, noise_std=0.01)        # Gaussian noise on actions (clipped to bounds)
obs, info = env.reset(seed=42)

noise_std can be a scalar (uniform across dimensions) or a per-dimension array. NoisyAction automatically clips noisy actions to the action space bounds.

Quick Environment Reference

Hugging Face Models and Datasets

Pre-trained model checkpoints and demonstration datasets are available on Hugging Face:

• Models: https://huggingface.co/kinder-bench
• Dataset: https://huggingface.co/datasets/kinder-bench/kinder-datasets

Acknowledgements

We thank the authors of following projects for open-sourcing their code, whose assets we utilized in KinDER:

• RoboCasa - for object assets in Dynamic3D envs
• MimicLabs - for scene assets in Dynamic3D envs

Contributing

General Guidelines

• All checks must pass before code is merged (see ./run_ci_checks.sh)
• All code goes through the pull request review process

Adding New Environments

Some new environment requests are in Issues. To add a new environment, please see the examples in src/kinder/env. Also consider:

• Environments are registered in src/kinder/__init__.py
• Each environment should have at least one demonstration (see scripts/collect_demos.py)
• After collecting a demonstration, create a video with scripts/generate_demo_video.py, which will be used in the autogenerated documentation