BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks

Yong Wang; Kecheng Zheng; Xing Zhu; Chao Xu; Yujun Shen; Shengze Cai

cs.ROOct 18, 2026

BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks

Authors: Yong Wang, Kecheng Zheng, Xing Zhu, Chao Xu, Yujun Shen, Shengze Cai

Organizations: Zhejiang University · Ant Group

Abstract

Bimanual manipulation with dexterous hands has attracted growing attention in research on robotic action models. However, few studies have developed benchmarks specifically tailored for dexterous manipulation, leaving the dexterous manipulation capabilities of existing models insufficiently characterized. In this paper, we present BiDex-100, a simulation benchmark for bimanual dexterous manipulation that combines broad task coverage, multi-dimensional evaluation, and high-DoF hand embodiments. Concretely, the benchmark comprises 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes, allowing systematic assessment of robustness. In addition, the benchmark supports two embodiments and provides 8,080 success-validated teleoperated demonstrations, totaling 6.25 million frames. We evaluate seven recent Vision-Language-Action models under a unified protocol, establishing baselines for capability-specific performance, robustness, and data efficiency. The website is available at https://BiDex-100.github.io/BiDex-100/.

Figures & tables

Benchmark	Biman.	DoF	Tasks	Progress	OOD	Demos
Meta-World [ 6 ]	×	–	50	✓	✓	×
RLBench [ 7 ]	×	–	100	×	×	synth.
LIBERO [ 8 ]	×	–	130	×	×	teleop.
CALVIN [ 9 ]	×	–	34	partial	✓	teleop
RoboCasa [ 10 ]	×	–	100	×	✓	teleop+synth.
ManiSkill2 [ 11 ]	✓	–	20	×	✓	synth.

TABLE I : Benchmark comparison. “DoF” = degrees of freedom of the single dexterous hand (– = parallel-jaw gripper, no multi-fingered dexterity); “Progress” = non-binary partial-credit scoring; “OOD” = built-in OOD protocol; “Demos” = released demonstrations ( teleop. = human-teleoperated, synth. = scripted/synthetic, “teleop+synth.” = both sources, “N/task” =

N

demonstrations per task, a bare count = total, ×= none).

Fig. 2 : Action Space Visualization. (a) Top-view log-density of end-effector positions across all 8,080 demonstrations; the two hands partition but substantially overlap the shared workspace, forcing bimanual coordination. (b) Per-joint usage of the bimanual 44 hand DoF, shown as median and 5–95% band; all five fingers engage, and thumb CMC and finger MCP joints carry the largest excursions.

Fig. 3 : Bimanual Collaborative Task. Fingertip trajectories of all ten fingers over the full motion reveal how contact-rich manipulation unfolds within a single episode. The right hand lifts the source cup above the recipient and tilts it to transfer the ice ball between the two cups, operating within centimeters while the fingers reconfigure among open-palm transport, precision pinch, and power grasp.

Fig. 4 : Dexterity Analysis. (a) Two-dimensional PCA of the bimanual 44-DoF posture over all 8,080 demonstrations; $k$ -means ( $k{=}5$ ) clusters, labeled by mean grasp-closure rank, span a continuous manifold from open palm to fist. (b) Cumulative PCA variance of the bimanual 44-DoF and single-hand 22-DoF posture: only eight components capture 80% of motion, so the high-DoF hand is governed by a compact synergy basis. (c) Per-joint loadings of the three leading grasp modes, grouped by finger for the right and left hands.

Fig. 5 : Overview of All Tasks. Full details on the tasks and data are available in Section XIII .

Fig. 6 : Matched ID / OOD Pairs. Rows show scene and layout changes, distractor injection, articulated-object rescaling, and left-right layout mirroring. The first row combines scene randomization with the task-specific OOD layout. The remaining rows illustrate optional transformations composed with the same OOD protocol. The task goal is unchanged, and each task, seed, and episode key is reproduced identically across models.

Panel of Figure 7 — Fig. 7 : Teleoperation hardware used for demonstration collection. (a) Joint-level tracking via an exoskeleton glove; (b) wrist-pose and hand-keypoint streaming from a head-mounted device.

Fig. 8 : Dataset Statistics. (a) Distribution of the 106 tasks and 8,080 demonstrations across the five capability axes. (b) Episode duration distributions, with the median marked for each axis. (c) Distribution of compositional depth measured by the number of subtasks per task. (d) The 14 most frequently used goal predicates among 42 predicate types across the 106 task specifications.

Fig. 9 : Object Asset Library. The 266 unique visible assets used by the benchmark, grouped semantically with the count shown for each category. Identical aliases are displayed once, the invisible assembly anchor is omitted, and all visible assets referenced by the task YAMLs are represented. Every asset is simulated with its mesh, mass, and friction properties.

Model	VLM backbone	Action expert	Chunk	Train
$\pi_{0.5}$ [ 4 ]	PaliGemma-2B	300M flow-matching	50	60k
GR00T-N1.7 [ 5 ]	Cosmos-R2-2B	DiT flow-matching	16	400k
InternVLA-A1.5 [ 36 ]	Qwen3.5-2B	1B flow-matching	50	400k
Galaxea G0.5 [ 37 ]	Qwen3.5-2B	flow-matching	32	400k
OpenDM-DM05 [ 38 ]	Gemma3-4B	rectified flow	50	120k
LingBot-VLA-v2 [ 39 ]	Qwen3-VL-4B	MoE flow (L1)	50	60k

TABLE II : Evaluated VLA baselines. “Chunk” = action chunk length; “Train” = fine-tuning optimizer steps, kept at each model’s official configuration. All models use 58-DoF joint control and are co-trained on one mixture containing demonstrations from all 106 tasks.

Model	Generalization	Instruction-Foll.	Limited-Demo.	Long-Horizon	Memory	Overall
LingBot-VLA-v2 [ 39 ]	61.35 / 75.80	35.90 / 66.57	66.15 / 76.14	38.20 / 66.78	26.00 / 43.20	52.21 / 70.22
$\pi_{0.5}$ [ 4 ]	55.45 / 72.70	36.60 / 64.48	66.23 / 75.63	34.00 / 64.51	33.00 / 50.16	50.40 / 68.97
GR00T-N1.7 [ 5 ]	46.55 / 65.75	24.60 / 59.28	61.38 / 73.97	32.80 / 62.32	20.80 / 36.10	42.32 / 63.43
OpenDM-DM05 [ 38 ]	33.05 / 53.18	1.80 / 23.53	35.69 / 49.11	3.20 / 17.45	12.60 / 26.13	23.06 / 40.66
Galaxea G0.5 [ 37 ]	24.90 / 44.51	5.70 / 36.33	39.62 / 53.20	5.20 / 25.10	10.60 / 28.22	21.68 / 41.73
Xiaomi XR1 [ 40 ]	6.40 / 24.48	0.90 / 21.06	7.15 / 21.71	0.00 / 15.63	0.60 / 7.00	4.40 / 20.67

TABLE III : In-Domain results. Success rate (%) / progress score per capability axis, averaged over 2 seeds

\times

25 episodes per task.

Model	Generalization	Instruction-Foll.	Limited-Demo.	Long-Horizon	Memory	Overall
LingBot-VLA-v2 [ 39 ]	1.60 / 10.76	0.80 / 19.90	3.46 / 10.06	0.00 / 7.15	0.80 / 6.18	1.68 / 11.54
$\pi_{0.5}$ [ 4 ]	2.40 / 13.19	2.40 / 23.96	6.31 / 13.54	0.40 / 10.02	3.80 / 10.70	3.30 / 14.77
GR00T-N1.7 [ 5 ]	4.65 / 17.59	3.10 / 30.25	10.23 / 21.34	2.60 / 19.96	3.80 / 12.59	5.45 / 20.65
OpenDM-DM05 [ 38 ]	0.85 / 10.89	0.10 / 17.04	3.92 / 10.97	0.00 / 5.72	1.80 / 7.28	1.47 / 11.24
Galaxea G0.5 [ 37 ]	0.20 / 4.40	0.20 / 15.41	2.62 / 5.75	0.00 / 3.66	0.20 / 1.57	0.77 / 6.47
Xiaomi XR1 [ 40 ]	0.05 / 2.26	0.00 / 14.95	0.00 / 1.47	0.00 / 2.89	0.00 / 2.18	0.02 / 4.51

TABLE IV : Out-of-Domain results. Success rate (%) / progress score per capability axis, averaged over 2 seeds

\times

25 episodes per task.

Fig. 10 : Overall performance of the evaluated VLA models. Strict success rate (%) followed by progress score in

[0,100]

, averaged over all 106 tasks and both evaluation domains. For every task, each model is evaluated with 2 fixed seeds and 25 episodes per seed in ID, and the same protocol in OOD. This gives 50 rollouts per domain, 100 rollouts per task, and 10,600 rollouts per model. The progress score distinguishes partial completion from complete failure.

Predicate	Semantics (tolerance $\epsilon$ unless noted)
Spatial placement
on	object center within a target’s horizontal radius and above it in height
stacked_on	object one layer above another within horizontal and vertical windows
on_table	object directly supported from below by the scene table
inside	object center within a circular horizontal radius and a vertical window
inside_contact	upward support contact from a container plus a minimum height

Table 14

Modality	Content	Dim
Observations
Arm end-effector pose (right, left)	xyz + quaternion, world or head-camera frame	14
Arm joint position / velocity	7 per arm	14 each
Hand joint position / velocity	22 per hand	44 each
Actions (one 58-dimensional control representation)
Arm component	joint targets or end-effector pose in world/head-camera frame	14

TABLE VII : Observation and action modalities per frame (25 Hz). The six video streams are RGB and depth from the head and the two wrist cameras.

Model	Train	Optimizer	Peak LR	Warmup	Min LR	GBS	Dtype	Chunk	Frozen modules
$\pi_{0.5}$	60k	AdamW (optax)	5e-5	100	5e-5	256	bf16	50	none (EMA 0.999)
GR00T-N1.7	400k	AdamW	1e-4	1500	$\sim$ 0	256	bf16	16	VLM + vision encoder
InternVLA-A1.5	400k	AdamW	5e-5	2000	5e-6	256	bf16	50	none
Galaxea G0.5	400k	AdamW	1e-5	1500	1e-6	256	bf16	32	none
OpenDM-DM05	120k	AdamW	2.5e-5	1000	2.5e-6	256	bf16	50	VLM token embeddings
LingBot-VLA-v2	60k	Muon (+AdamW)	1e-4	0	5e-5	256	fp32	50	none

TABLE VIII : Training hyperparameters and selected checkpoints of the seven evaluated VLAs (cosine-family schedules, grad-clip 1.0). GBS = global batch size.

Bimanual Robotic Manipulation

Bimanual robotic manipulation coordinates two robot arms to handle objects, tools, and tasks that require collaboration.

See all 178 papers

Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands
Sep 14, 2026Zhenjie Yang, Yideng Zhang, Dongjie Zhang +21Dexterous Manipulation Bimanual Robotic Manipulation
The University of Hong Kong · Shanghai Jiao Tong University · Fudan University +4
DexVerse: A Modular Benchmark for Multi-Task, Multi-Embodiment Dexterous Manipulation
Jul 9, 2026Yunchao Yao, Zhuxiu Xu, Tianqi Zhang +12Dexterous Manipulation Long-Horizon Robotic Manipulation
UNC-Chapel Hill · The University of Hong Kong · UC Berkeley
DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo
May 15, 2026Hanwen Wang, Weizhi Zhao, Xiangyu Wang +11Dexterous Manipulation Long-Horizon Robotic Manipulation
NLPR & MAIS, CASIA · SJTU · MBZUAI +3

Sep 14, 2026J/K move · Enter open · S save

Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands

Zhenjie Yang, Yideng Zhang, Dongjie Zhang, Chenyu Jiang+20

The University of Hong Kong · Shanghai Jiao Tong University · Fudan University+4

Dexterous Manipulation Bimanual Robotic Manipulation Robotics Simulation Robotic Manipulation Robot Manipulation Benchmarks Tactile Sensing for Robotic Manipulation

Open paper AskPDF arXiv

Tactile sensing provides contact information that can be difficult to infer from vision alone, but tactile hardware for dexterous hands has not converged to a common design. Dexterous hands differ in finger structure, contact surfaces, and sensor layouts, while simulated tactile signals still differ from measurements produced by physical sensors. These factors make it difficult to study visuo-tactile manipulation across diverse dexterous hands within a consistent experimental setting. We present Bench2Dex, a simulation benchmark for visuo-tactile bimanual manipulation across 12 dexterous hands. We adapt existing robot models with a shared simulated tactile interface that converts local contact geometry into image-like tactile observations. The interface provides a consistent observation format across different hand morphologies without attempting to reproduce the output of a specific physical tactile sensor. Bench2Dex includes 26 bimanual manipulation tasks that involve tool use, articulated-object interaction, and multi-stage manipulation, together with about 1.3K human-teleoperated demonstrations. The benchmark provides synchronized visual, tactile, proprioceptive, action, and object-state observations, together with executable task metrics. For robustness, we group seven perturbation types into invariance axis, where the correct action does not change, and equivariance axis, where the correct action changes together with the perturbation. We evaluate ACT, Diffusion Policy, pi0.5, and GR00T N1.5 on Bench2Dex and report their performance and failure modes. Bench2Dex is meant as a platform for studying visuo-tactile learning across dexterous hands. It does not assume that simulated tactile observations can replace real tactile sensing; it offers a shared setting for algorithm development while tactile hardware and simulation models are still evolving.

Abstract

Bimanual Robotic Manipulation

Explore similar work

Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands

DexVerse: A Modular Benchmark for Multi-Task, Multi-Embodiment Dexterous Manipulation

DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo