Dexsuite

Abstract

Dexterous multi-fingered hands promise far richer manipulation skills than simple parallel grippers, yet most existing benchmarks and datasets still target low-DoF end-effectors. In an era where increasingly advanced physical hands are being developed at high speed, we lack a unified, low-cost simulation and benchmarking environment to study dexterous control. We introduce Dexsuite , a modular framework that standardizes observations, action spaces, and evaluation protocols for multi-fingered hands while remaining easily extensible to new robots, environments, and tasks. Alongside the framework, we release a curated dataset of over 150k frames of multi-modal hand interaction, collected with a Manus glove on a suite of manipulation tasks, providing high-fidelity supervision of finger motion and contact-rich interactions. DexSuite also offers benchmark tasks and baselines spanning imitation learning, reinforcement learning, and diffusion policies, enabling fair comparison across algorithm families and hand representations, and providing a common testbed for systematic study of robotic dexterity.

DexSuite Framework

DexSuite is a unified simulation framework and benchmark for dexterous manipulation. It separates a Base Environment from modular Tasks and uses robot abstractions to pair any manipulator with any gripper or multi‑finger hand in single or bimanual settings across rigid, articulated, and deformable objects.

Dataset

We provide a 150k frame dexterity dataset spanning 15 environments with four vision modalities: RGB, depth, segmentation, and surface normals. The dataset is stored in the LeRobotDataset format with episode trajectories, synchronized timestamps, and metadata for single and bimanual variants. Loaders are included to stream batches for imitation and offline RL. We also provide utilities to convert to Robomimic or RLDS dataset.

We provide multiple teleoperation devices, ranging from simple keyboard to specialized hardware such as Manus Gloves for Hand data collection.

We also provide native integration of retargeting tools both direct qpos to qpos and more advanced retargeting.

Native Vive Tracker/Controller integration for Pose control.

Manus glove path from real to simulator using Geometric retargeting, producing strong teleoperation dexterous hands.

Robots

DexSuite supports modular manipulators with swappable hands and integrated robots with native grippers. The catalog covers configurations from Panda with Allegro to dual‑arm presets in one place.

Franka Panda

Franka FR3

Kinova Gen3

Inspire Manipulator

KUKA LBR iiwa

UR5e

UR10

SO‑100

WX250s

Link 6

Z1

Sawyer

Lite6

Ability Hand

Allegro Hand

Barrett Hand

DClaw

Inspire Hand

LEAP Hand

Robotiq 2F-85

Schunk Hand

Shadow Hand

UMI Hand

Mix and match manipulators and grippers. Not all hands are compatible with all manipulators.

Any Robot Setup

Our environment initialization function allows you to make any environment, with custom Simulator, Robot, and Cameras without ever touching Dexsuite code. We also provide a web based environemnt maker that allows user to make any config, copy-paste the code and run it.

Droid

Aloha

Environment Visualizer

Select an environment to preview it before the table.

Pick and Place Pan

Available Environments

The suite provides single-arm and bimanual environments across rigid, articulated, and deformable object tasks.

Environment	Mode	Type
Reach	Single-arm	Rigid
Push	Single-arm	Rigid
Lift	Single-arm	Rigid
Pick Place	Single-arm	Rigid
Pick Place Fruit	Single-arm	Rigid
Pick Place Mug	Single-arm	Rigid
Pick Place Pan	Single-arm	Rigid
Stack	Single-arm	Rigid
Make Coffee	Single-arm	Rigid
Drill to Point	Single-arm	Rigid
Tool Hang	Single-arm	Rigid
Bimanual Reach	Bimanual	Rigid
Bimanual Push	Bimanual	Rigid
Bimanual Lift	Bimanual	Rigid
Bimanual Pick Place	Bimanual	Rigid
Bimanual Pick Place Mug	Bimanual	Rigid
Bimanual Pick Place Pan	Bimanual	Rigid
Bimanual Pick Place Pot	Bimanual	Rigid
Bimanual Stack	Bimanual	Rigid
Bimanual Stack Mid Air	Bimanual	Rigid
Bimanual Make Coffee	Bimanual	Rigid
Bimanual Drill to Point	Bimanual	Rigid
Bimanual Fold Glasses	Bimanual	Articulated
Open Jar	Single-arm	Articulated
Close Jar	Single-arm	Articulated
Open Faucet	Single-arm	Articulated
Close Faucet	Single-arm	Articulated
Cable Routing	Single-arm	Deformable
Cut Butter	Single-arm	Deformable
MPM Sponge Drop	Single-arm	Deformable
Pour Water	Single-arm	Deformable
Spread Butter	Single-arm	Deformable

Benchmarks

Imitation Learning

Teleoperation datasets power the imitation learning suite. We provide glove‑conditioned demonstrations, dataset utilities, and offline evaluation scripts so you can train supervised policies before running closed‑loop simulation.

Robotiq

Task	BC-G	BC-GMM	Transformer	Diffusion
Reach	91 ± 4	82 ± 4	99 ± 1	71 ± 1
Lift	71 ± 2	71 ± 5	87 ± 3	29 ± 1
Push	43 ± 5	51 ± 2	45 ± 6	57 ± 8
Pick&Place	61 ± 6	63 ± 5	69 ± 4	66 ± 7
Stack	2 ± 1	0 ± 0	5 ± 2	1 ± 1

Allegro

Task	BC-G	BC-GMM	Transformer	Diffusion
Reach	63 ± 10	71 ± 10	79 ± 4	59 ± 1
Lift	73 ± 8	40 ± 4	65 ± 4	69 ± 3
Push	40 ± 5	24 ± 7	58 ± 5	55 ± 8
Pick&Place	18 ± 5	29 ± 8	35 ± 1	19 ± 4
Stack	0 ± 0	0 ± 0	0 ± 0	1 ± 1
Rack Mug	37 ± 7	55 ± 4	53 ± 3	38 ± 6
Move Pan	4 ± 2	68 ± 11	60 ± 7	50 ± 8
DrilltoPoint	38 ± 7	62 ± 3	59 ± 4	83 ± 9
Place Fruits	0 ± 0	0 ± 0	2 ± 0	9 ± 3
Bimanual Stack Mid Air	0 ± 0	0 ± 0	0 ± 0	0 ± 0

Reinforcement Learning

DexSuite provides vectorized rollouts and a unified reward API. PPO and SAC recipes are tuned for rigid and deformable manipulation.

RL‑evaluated Environments

Environment	PPO	SAC
Reach
Lift
Stack

BibTeX

                        
  @article{articlename,
    author    = {Anonymous},
    title     = {Dexsuite.},
    journal   = {TBD},
    year      = {2025},
  }