dexsuite 0.1.0

Dexterous manipulation. Simulated, teleoperated, benchmarked.
All in one framework.

Website · Documentation · Datasets

DexSuite brings the robot, the tasks, teleoperation, tactile sensing, the data and the evaluation into one stack, on the Genesis simulator. A task is written once and runs with any arm and any hand, from a parallel-jaw gripper to a 24-joint hand.

Install

pip install dexsuite                # every arm, hand and task
pip install "dexsuite[all]"         # plus SpaceMouse, Vive and Manus teleoperation and the LeRobot dataset tools

Python 3.10 or newer and a CUDA GPU (the CPU works for a quick look). The first environment on a machine compiles the Genesis kernels, which takes about a minute. On a machine without a display, export PYOPENGL_PLATFORM=egl.

Quickstart

import dexsuite as ds

env = ds.make("lift", manipulator="franka", gripper="allegro")
obs, info = env.reset(seed=0)

for _ in range(300):
    action = env.action_space.sample()       # 6 values move the hand, 16 set the Allegro joints
    obs, reward, terminated, truncated, info = env.step(action)

Every environment is a Gymnasium environment with batched PyTorch tensors: n_envs=64 builds 64 copies in one simulator. The documentation covers the tasks, the robots, the actions and observations, teleoperation and the benchmark datasets on Hugging Face.

Citation

@article{dexsuite2026,
  author  = {Anonymous},
  title   = {DexSuite: An All-in-One Framework for Dexterous Manipulation Across Arms, Hands, and Tasks},
  year    = {2026},
}