MINT-SJTU/Evo-DexBench-dataset

Evo-DexBench: Benchmarking Generalization across Related Tasks for Dexterous Manipulation

30 tasks | 6 functional families | 1,248 successful demonstrations | 3 scene layouts per task | HDF5 and LeRobot v3

Evo-DexBench is a benchmark and demonstration dataset for studying whether dexterous-manipulation policies can transfer functional interaction knowledge from related Core tasks to a new Gen task with only a few target demonstrations. Each task family contains four Core tasks and one Gen task. The Core–Gen design controls source–target functional relatedness and target-task exposure, making it possible to study transfer separately from task memorization.

The dataset contains multimodal demonstrations collected in simulation with Franka Panda arms and simulated Shadow Hands. Single-arm tasks use one Panda–Shadow Hand unit; bimanual tasks use two. It is intended for research on dexterous robot learning, few-shot task adaptation, cross-task generalization, tactile-informed manipulation, and robustness to scene variation.

Highlights

• Core–Gen transfer protocol: four related Core tasks provide source experience for a held-in-family Gen target.
• Six functional families: button-driven manipulation, rotational manipulation, access-and-transfer, structured configuration, assembly, and articulated configuration control.
• Few target demonstrations: each Core task has 50 successful demonstrations and each Gen task has 8, for 1,248 demonstrations total.
• Scene diversity: each task appears in three layouts across Studio and Kitchen environments.
• Multimodal data: synchronized multi-view RGB and depth, robot state and action, task instruction and stage information, and tactile measurements.
• Two data formats: canonical HDF5 for the full multimodal record and LeRobot v3 for policy training.

Dataset Summary

Property	Value
Dataset	MINT-SJTU/Evo-DexBench-dataset
Tasks	30
Functional families	6
Core / Gen tasks	24 / 6
Successful demonstrations	1,248
Demonstrations per Core task	50
Demonstrations per Gen task	8
Scene layouts	3 per task (90 task-layout combinations)
Environments	Studio and Kitchen
Robot embodiment	Franka Panda + simulated Shadow Hand; bimanual tasks use two Panda–Shadow Hand units
Tactile sensing	123 virtual taxels per hand; 246 channels for dual-hand observations
Data formats	Canonical HDF5; LeRobot v3
Control / observation rate	30 Hz; simulator physics at 120 Hz
Conventional train/validation/test split	Not specified; Core and Gen define the transfer roles

Task Families

Family	Interaction	Core tasks	Gen task
Button-driven manipulation	Displace a localized control to change a device state	LampButton, WallSwitch, PhoneHomeButton, ToasterSlider	DisplayPowerButton
Rotational manipulation	Apply directed rotation to a functional component	FaucetHandle, MouseWheel, CoffeeKnob, KettleLidTab	DishwasherKnob
Access-and-transfer	Open an access mechanism, move an object, and restore the mechanism state	CupFromCabinet, MicrowaveBowl, TennisBallTransfer, BananaFromDrawer	CandleToCabinet
Structured configuration	Build an arrangement subject to ordering or stability constraints	TablewarePlacement, HammerBlockAssembly, SandwichAssembly, RingTowerAssembly	BoxPacking
Assembly	Align complementary objects, often with two hands, for insertion or fitting	PenCapping, CupSleeveFit, BottleFunnelInsertion, PlugInsertion	ScissorsSheathing
Articulated configuration control	Stabilize one part while changing another part's relative configuration	LaptopClosing, StaplerPaperMark, ScissorsOpening, BoxLidOpening	GlassesFolding

The first three families use a single arm and hand; the last three use a dual-arm setup. Access-and-transfer and structured-configuration tasks are long-horizon tasks.

Data Contents

The policy-independent demonstration record is sampled at 30 Hz and includes synchronized RGB and depth observations, robot state, the action applied by the demonstrator, a natural-language task instruction, task-stage state, and tactile measurements. The simulator advances at 120 Hz.

The benchmark's common policy interface defines the following compact state and action dimensions:

Setup	Released LeRobot observation.state	Paper's compact policy state	action
Single arm and hand	111 dimensions	31 dimensions: 7-D end-effector pose + 24 hand-joint positions	30 dimensions: 6-D end-effector pose increment + 24 absolute hand-joint targets
Dual arm and hands	222 dimensions	62 dimensions	60 dimensions

The visual input uses context, interaction, and wrist views. The released LeRobot files also contain aligned depth observations for these cameras. Single-arm policy inputs use the context, interaction, and right-wrist RGB views; bimanual inputs use context, left-wrist, and right-wrist views. Each Shadow Hand provides a 123-dimensional non-negative normal-force signal in newtons. For bimanual tasks, the left-hand taxel layout mirrors the right-hand layout.

The released LeRobot observation.state is 111-D for single-arm tasks and 222-D for bimanual tasks. The compact 31-D/62-D state above is the paper's common model interface after preprocessing; the 30-D/60-D action dimensions match the released action feature.

The full tactile measurements are exposed by the benchmark interface. Individual policies may use a subset; for example, the paper's KineDex experiments consume the finger taxels and do not pass palm taxels to that model.

Field Reference

The release provides two full-data schemas: canonical per-episode HDF5 and task-wise LeRobot v3.

Canonical HDF5

In the paths below, N is the number of observations in an episode. Actions and transitions normally have N - 1 rows because each one links adjacent observations. {view} is a camera role present for the task, and {side} is an active manipulator (right for single-arm tasks; left and right for bimanual tasks).

HDF5 path	Shape / type	Description
observations/images/{view}	uint8 [N, 224, 224, 3]	Synchronized RGB images from context, interaction, and wrist cameras.
observations/depth/{view}	int16 [N, 224, 224]	Depth map aligned with the corresponding camera; the stored encoding is environment-defined, so values should not be assumed to be meters.
observations/state/{side}.joint_position, .joint_velocity	float32 [N, 31] each	Panda arm (7) and Shadow Hand (24) joint positions or velocities.
observations/state/{side}.end_effector_pose, .palm_pose	float32 [N, 7] each	End-effector or palm position and orientation.
observations/state/{side}.fingertip_poses	float32 [N, 5, 7]	Position and orientation for the five fingertips.
observations/taxel/{side}	float32 [N, 123]	Non-negative distributed tactile normal-force readings in newtons.
actions/native/*	float32 [N - 1, 30] per active controller	Simulator-native action vector. Single-arm files use value; bimanual files store separate left and right vectors.
actions/semantic/{side}/end_effector_delta_pose	float32 [N - 1, 6]	End-effector translation delta in meters in the robot-root frame, plus XYZ Euler rotation delta in radians in the root-aligned body-rotation frame.
actions/semantic/{side}/hand_joint_position	float32 [N - 1, 24]	Absolute Shadow Hand joint-position targets in radians.
observations/time_s, actions/time_s, transitions/*	Timestamp / ID / JSON fields	Observation and action timing, transition links, held-action flags, and transition info, reward, termination, and truncation records.
meta/*, integrity/*	JSON metadata / SHA-256	Episode instruction and identifiers, required cameras and manipulators, action and sensor specifications, timing, completion status, and chunk integrity hashes.

HDF5 episodes may also include diagnostics/quest/*, which stores raw Quest hand-tracking and retargeting diagnostics; these are supplementary to the policy-facing observations and actions.

LeRobot v3

The task-wise LeRobot export uses these principal features:

Feature	Type / shape	Description
observation.state	float32 [111] single-arm; float32 [222] bimanual	Full robot state, including arm/hand joint positions and velocities plus end-effector, palm, and fingertip poses.
action	float32 [30] single-arm; float32 [60] bimanual	Released per-arm action interface; each arm contributes a 6-D end-effector delta and 24 hand-joint targets.
observation.images.{view}	Video, 224 × 224 × 3, 30 Hz	RGB context, interaction, and wrist-camera streams. Single-arm tasks use the right wrist; bimanual tasks include left and right wrists.
observation.depth.{view}	float32 [224, 224]	Depth maps aligned with the corresponding image views.
observation.tactile.{side}	float32 [123]	Per-hand tactile normal-force measurements in newtons; bimanual tasks include both sides.
next.reward, next.terminated, next.truncated	float32 [1], bool [1], bool [1]	Reward and episode-ending flags for the transition.
timestamp, frame_index, episode_index, index, task_index	Scalar numeric fields	Time and row/episode/task indexing. task_index maps to the language instruction in meta/tasks.parquet; episode lengths and video ranges are recorded under meta/episodes/.

The released observation.state is larger than the paper's compact 31-D/62-D policy state because it retains additional robot and fingertip state. The compact policy state is derived during preprocessing.

Repository Layout

Each HDF5 file stores one demonstration episode. The LeRobot v3 export is organized as one dataset directory per task:

Evo-DexBench-dataset/
├── h5/<family>/<task>/*.h5
└── lerobot/<family>/<task>/
    ├── meta/
    │   ├── info.json
    │   ├── tasks.parquet
    │   └── episodes/chunk-000/file-000.parquet
    ├── data/chunk-000/file-*.parquet
    └── videos/observation.images.<view>/chunk-000/file-*.mp4

The paper does not specify a fixed train/validation/test split. For transfer experiments, demonstrations are organized by task role: Core data are the related source tasks, and Gen data are few-shot target demonstrations. The paper's baseline evaluation covers 15 tasks from the button-driven, rotational-manipulation, and articulated-configuration-control families; the full benchmark contains 30 tasks.

Collection

Demonstrations were collected in simulation through a VR teleoperation pipeline using Quest 3 WebXR hand tracking. The operator's wrist motion is mapped to the Panda arm end-effector, and hand keypoints are retargeted to Shadow Hand joint targets. Each task is collected across three scene layouts selected at random.

Scene assets include YCB objects, PartNet-Mobility articulated objects, RoboCasa kitchen scenes, and custom assets. Context and interaction cameras are calibrated per task; a wrist camera provides close-up interaction detail.

Tactile sensing uses a distributed virtual-tactile representation. Each hand has 123 taxels reporting non-negative normal force in newtons. In the simulation, contact impulses are accumulated at the 120 Hz physics rate and converted to force observations aligned with the 30 Hz policy/control rate.

Evaluation Protocol

The paper evaluates policies using 100 episodes per task: 20 nominal episodes and 20 episodes for each of four single-factor variation axes: camera, lighting, physical properties, and spatial configuration. Compared policies share a registered reset schedule, and model weights remain frozen during evaluation.

Reported metrics are task success rate and normalized stage score. Aggregate results are macro-averaged across tasks so each task contributes equally. A stage score represents the highest task-specific progress stage reached during an episode.

Intended Uses

Evo-DexBench is intended for research on:

• Few-shot transfer from related dexterous-manipulation tasks
• Learning and evaluating reusable functional interactions
• Vision-language-action and visuomotor policy learning
• Tactile-informed manipulation
• Robustness to camera, lighting, physical, and spatial variation
• Benchmarking single-arm and bimanual dexterous policies

It is a research benchmark and is not a substitute for task-specific validation before deployment on physical robots.

Acknowledgements

The benchmark uses assets and software from YCB, PartNet-Mobility, RoboCasa, SAPIEN/ManiSkill, and related upstream projects and datasets. Please consult the project repository and the respective upstream terms for attribution and use conditions.

Citation

If you use Evo-DexBench, please cite the benchmark paper:

@misc{cheng2026evodexbench,
  title        = {Evo-DexBench: Benchmarking Generalization across Related Tasks for Dexterous Manipulation},
  author       = {Cheng, Bing and Mao, Yiran and Ye, Zewei and Jin, Yifan and Chen, Xinyi and Tu, Ruifeng and Wang, Shuhuai and Lin, Tao and Zhou, Zhimu and Du, Yuxin and Zhong, Fangwei and Yan, Junchi and Zhao, Bo},
  year         = {2026},
  url          = {https://arxiv.org/abs/2610.21180}
}

• Hugging Face dataset

---