Literature
Display
← All papers

Dexora: Open-source VLA for High-DoF Bimanual Dexterity

Zongzheng Zhang · Jingrui Pang · Zhuo Yang · Kun Li · Minwen Liao · Saining Zhang · Guoxuan Chi · Jinbang Guo · Huan-ang Gao · Modi Shi · Dongyun Ge · Yao Mu · Jiayuan Gu · Rui Chen · Hao Dong · Huazhe Xu · Li Yi · Yixin Zhu · Hang Zhao · Pengwei Wang · Shanghang Zhang · Guocai Yao · Jianyu Chen · Hongyang Li · Hao Zhao

· ICRA 2026 ↗

Robotics8 min text · CondensedOriginal paper ↗
Contents

At a glance

Dexora combines a 36-DoF dual-arm, dual-hand platform, matched simulation/real demonstrations and a discriminator that weights diffusion-policy training. Its main results are 89.6% mean basic-task success and 66.7% dexterous-task success, compared with the strongest listed baseline’s 82.1% and 51.7%. These are results on small, specified test suites; robot transfers require task demonstrations, and cap twisting remains difficult.

I Introduction

The work targets coordinated arm motion and fine finger control in one policy. Two 6-DoF arms and two 12-DoF hands support handovers, constrained retrieval and torsional tasks that simpler tested configurations handle poorly. An exoskeleton captures arm motion, Vision Pro tracks fingers, and the same interface controls a physical robot and a MuJoCo twin.

The central recipe is simulation pretraining followed by real-data specialization with quality weighting. The claim of a universal high-to-low-dimensional controller is a hypothesis motivated by adapted examples, not a theorem about arbitrary robot kinematics. This edition follows all eight pages of the 18 May 2026 v1, retaining all eleven figures, three tables, eight numbered equations and 44 references. Original resources are linked at the project site.

Complete original PDF figure S1.F1
Fig. 1: Dexora overview. (a) Motivation: Three illustrative contrasts highlight the need for dual-arm, dual-hand dexterous VLA: piston insertion (requires two arms), book retrieval from a packed shelf (hands with fingers succeed where grippers fail), and bottle opening (12-DoF fingers with lateral swing outperform 6-DoF). (b) Dataset (§III-B): We pretrain on 100K simulated bimanual-hand trajectories and post-train on 10K real demonstrations, all collected with our dual-arm, dual-hand platform. (c) Architecture (§III-C): A trained discriminator scores dataset demonstration quality and guides training, driving the diffusion-transformer policy to prioritize high-quality trajectories while down-weighting low-quality ones. (d) Performance (§IV-B): Dexora achieves consistently higher average success rates on both basic (Pick-and-Place, Assemble/Disassemble, Articulated Object) and dexterous benchmarks compared to state-of-the-art VLA models. (e) Embodiment generalization (§IV-C): The same policy transfers across single-arm gripper, dual-arm grippers, and single-arm low-DoF hand without re-architecting the model.
Complete original PDF figure S1.F2
Fig. 2: Comparison of embodiment coverage. Prior works cover either single-arm or low-DoF dual-arm settings. Dexora is the first system positioned in the dual-arm, high-DoF dexterous region, while also generalizing across simpler embodiments without re-architecture.

II Related Work

II-A Teleoperation System

Teleoperation can use leader–follower rigs, headsets, vision retargeting, exoskeletons or handheld controls. The hybrid setup combines joint-space arm capture, as in AirExo, with glove-free finger tracking related to Bunny-VisionPro. Matching the interface across physical and simulated embodiments makes collected commands consistent; the paper does not isolate interface latency or operator fatigue in a quantitative comparison.

II-B Dexterous Manipulation

Prior work separates grasp synthesis from policy learning. Diffusion, flows, latent generative models and optimization produce grasps; reinforcement and imitation learning turn them into control. DexMimicGen scales demonstrations through retargeting, while EgoVLA uses human video. Dexora’s emphasis is language-conditioned coordination of both arms and high-dimensional hands, rather than grasp generation alone.

II-C Vision-Language-Action (VLA) Model

RT-2, OpenVLA and GraspVLA motivate generalist robot policies; π₀, RDT and GR00T N1 extend the architectural and embodiment range. The source’s coverage diagram is a selected comparison, not an exhaustive proof of priority. Dexora learns synchronized arm–hand action chunks, then fine-tunes for simpler robots.

Complete original PDF figure S2.F3
Fig. 3: Hardware and teleoperation system. (a) Hybrid teleoperation interface and 12-DoF XHAND. (b)-(c) The operator teleoperates the physical robot and its MujoCo digital twin, so apple→\rightarrowplate demonstrations are collected in real and simulation under the same interface, thereby reducing the sim-to-real gap.

III Dexora

The system links four components: a shared hardware/simulation interface, task-diverse demonstrations, a conditional diffusion policy and an offline quality scorer. The scorer influences training weights and is removed at deployment.

III-A Dual-Arm Dual-hand System

Each AIRBOT arm has six joints and each XHAND twelve independently actuated joints; thumb/index lateral motion supports reorientation and twisting. The backpack maps measured arm angles directly into robot joint commands. Vision Pro finger skeletons are calibrated and retargeted with joint limits. Four RGB views and all 36 joint states are synchronized at 20 Hz.

The MuJoCo twin shares kinematics, controllers and teleoperation drivers with the physical system. This reduces an interface mismatch but does not establish equality of contact physics or sim-to-real performance. The authors attribute smoother arm commands to avoiding vision-only inverse-kinematics jitter; that attribution is not separately benchmarked.

III-B Dataset Construction

Simulation. Qwen2.5-VL selects manipulable Objaverse objects and proposes physical parameters, layouts and success criteria. For each of 200 tasks, three to five demonstrations seed DexMimicGen-style retargeting to randomized scenes, producing 500 trajectories per task: 100,000 total, with 6.5 million reported frames. Figure 4 shows 297 simulation objects in 30 categories.

Real data. Two hundred tasks with fifty demonstrations each yield 10,000 episodes, including tool use and a 20% dexterous-task share. Figure 4 lists 347 objects in seventeen categories. Data are converted to the stated LIBERO-2.1 format. The methods report 40.5 hours and 2.92 million real frames; the introduction instead says 177.5 hours and 3.2 million. The simulation’s 361 video-hours also has an unclear per-view counting convention at the stated 20 Hz. These totals should not be treated as a single reconciled duration account.

Complete original PDF figure S3.F4
Fig. 4: Dataset demonstration. (a) Simulation objects subset: our simulator includes 297 objects across 30 categories. (b) Real-world objects (347 objects, 17 categories), covering both basic and dexterous use cases. (c) Per-family task distribution in simulation vs. real. The simulation data only includes basic tasks, while the real-world set shifts weight toward dexterity (20%). (d) Trajectory counts per family and embodiment (sim/real; single-/dual-hand).

III-C Framework

Quality labels. Joint states are min–max normalized, then centered finite differences produce velocity, acceleration and jerk. Equations 2–3 aggregate acceleration and jerk over time and dimensions. The intersection of the lowest 20% by each metric retains about 18% of episodes; replayed task success without collisions reduces this to about 15% positives. A movement-coverage guard is mentioned to reject uninformative stationary behavior, but its threshold is not specified.

Discriminator. Uniformly sampled subclips contain state, action chunks and a frozen-policy compatibility score, conditioned on language and images. Equations 4–5 use standardized negative denoising residual energy. This is a relative compatibility proxy, not an evaluated action log-likelihood. The printed formula omits explicit diffusion-step/noisy-action notation, and the number of sampled clips/steps and clip-to-episode aggregation are unspecified. Transformer features are pooled and passed through an MLP/sigmoid quality head.

Policy. A decoder-only diffusion Transformer combines state, noisy action and timestep tokens with alternating T5 language and SigLIP image conditioning. It predicts action noise, using DDPM training and DPM-Solver++ inference. The policy emits synchronized joint-command chunks; the discriminator is unnecessary during execution.

Complete original PDF figure S3.F5
Fig. 5: Dexora framework. (a) Data filtering: From the real-world dataset we pre-screen demonstrations by kinematic smoothness (low acceleration and jerk), then replay them for post-validation and keep the clips that complete the task without collisions, forming a high-quality subset. (b) Discriminator training: With the pretrained diffusion–transformer policy frozen, we compute a log-π\pi proxy for each clip and train a discriminator that, conditioned on observations and language, outputs a quality score d⁡(Ct)∈(0,1]d(C_{t})\in(0,1]. (c) Data-quality-aware post-training: During post-training, the score d⁡(Ct)d(C_{t}) is converted to weights wiw_{i} and used in the diffusion loss ℒπ\mathcal{L}_{\pi}. At inference time, only the policy is used.
Equations 1–5
vt\displaystyle v_{t} =st+1−st−12​Δ​t,at\displaystyle=\frac{s_{t+1}-s_{t-1}}{2\Delta t},\quad a_{t} =vt+1−vt−12​Δ​t,jt\displaystyle=\frac{v_{t+1}-v_{t-1}}{2\Delta t},\quad j_{t} =at+1−at−12​Δ​t,\displaystyle=\frac{a_{t+1}-a_{t-1}}{2\Delta t},\quad (1)
Aep​(τ)\displaystyle A_{\text{ep}}(\tau) =1(T−6)​D​∑t=4T−3∑k=1Dat,k2,\displaystyle=\sqrt{\frac{1}{(T-6)D}\sum_{t=4}^{T-3}\sum_{k=1}^{D}a_{t,k}^{2}}, (2)
Jep​(τ)\displaystyle J_{\text{ep}}(\tau) =1(T−6)​D​∑t=4T−3∑k=1Djt,k2.\displaystyle=\sqrt{\frac{1}{(T-6)D}\sum_{t=4}^{T-3}\sum_{k=1}^{D}j_{t,k}^{2}}. (3)
Et=1|𝒮|​L∑s∈𝒮∑τ=tt+L−1‖εθ(𝐨τ,ℓ,𝐚τ:τ+L−1,sτ)−ε‖22,E_{t}=\frac{1}{|\mathcal{S}|\,L}\sum_{s\in\mathcal{S}}\sum_{\tau=t}^{t+L-1}\left\|\varepsilon_{\theta}\!\left(\mathbf{o}_{\tau},\,\ell,\,\mathbf{a}_{\tau:\tau+L-1},\,s_{\tau}\right)-\varepsilon\right\|_{2}^{2}, (4)
log⁡π^t=−zscore⁡(Et)=−Et−Mean​(E)Var​(E)+ε,\widehat{\log\pi}_{t}=-\,\mathrm{zscore}\!\left(E_{t}\right)=-\frac{E_{t}-\text{Mean}(E)}{\sqrt{\text{Var}(E)+\varepsilon}}, (5)
Complete original PDF figure S3.F6
Fig. 6: Basic tasks suite. (a) Pick and Place (5 tasks). (b) Assemble/Disassemble (5 tasks). (c) Articulated Objects (2 tasks).
Equation 6
πθ(st,𝐨t,ℓ)=𝐚^t:t+L−1.\pi_{\theta}(s_{t},\ \mathbf{o}_{t},\ \ell)=\widehat{\mathbf{a}}_{t:t+L-1}. (6)

III-D Data-quality-aware Training Recipe

First pretrain the policy on simulation; then freeze it to compute compatibility features and train the discriminator; finally post-train the policy on all real demonstrations using precomputed weights. Equation 7 assigns replay-validated positives a weighted binary-cross-entropy target of one and treats the remaining, unlabelled pool as negative, with η=0.5.

Although called positive–unlabelled learning, the displayed objective has no correction for genuine high-quality examples remaining in that negative pool. Scores are clipped to [0.1, 0.9], then mapped to weights following discriminator-weighted behavior cloning. The mapping, claimed calibration procedure and warm-up duration are not reproduced. Equation 8 preserves the weighted denoising loss; the scores should be read as practical reweighting signals rather than certified quality probabilities.

Equations 7–8
ℒD=η​𝔼τ∈𝒮high​[−log⁡d⁡(τ)]⏟positive BCE→ 1+𝔼τ∈𝒰​[−log⁡(1−d⁡(τ))]⏟unlabeled as negative→ 0,\mathcal{L}_{D}=\eta\,\underbrace{\mathbb{E}_{\tau\in\mathcal{S}_{\mathrm{high}}}\!\big[-\log d(\tau)\big]}_{\text{positive BCE}\;\to\;1}\;+\;\underbrace{\mathbb{E}_{\tau\in\mathcal{U}}\!\big[-\log(1-d(\tau))\big]}_{\text{unlabeled as negative}\;\to\;0}, (7)
ℒπ=∑i=1Lwi​‖εθ​(⋅)−ε‖22,\mathcal{L}_{\pi}=\sum_{i=1}^{L}w_{i}\;\big\|\varepsilon_{\theta}(\cdot)-\varepsilon\big\|_{2}^{2}, (8)

IV EXPERIMENT

Experiments cover native-platform basic and dexterous skills, six visual/environmental shifts of one task, adapted embodiments, and separate data-composition/discriminator ablations. Each reported task evaluation uses twenty rollouts, so individual-task outcomes change in five-percentage-point steps. Training-seed uncertainty and confidence intervals are not reported.

IV-A Experimental Setup and Baselines

The policy has 28 layers, width 1024 and sixteen attention heads. The 30-million-parameter discriminator has twelve layers, width 512 and eight heads. Policy pretraining runs 100,000 steps and discriminator training 10,000, using AdamW, eight A100 GPUs and total batch size 64.

Diffusion Policy directly predicts the 36-dimensional action. π₀ uses a two-layer projector factored into left/right arm/hand groups to expand its native outputs; GR00T N1 is the strongest baseline. Chunk length is 32, with matched cameras, views and control frequency. Baselines receive 100 demonstrations per task and 50,000 fine-tuning steps; the protocol uses LoRA on four L20 GPUs and RTX 4090 inference. Basic Dexora results are described as zero-shot, whereas dexterous and cross-robot experiments explicitly fine-tune with demonstrations. Different prior training and hand capacities remain comparison factors.

IV-B Evaluation Results in Real World

Basic tasks. Table I contains five pick/place, five assembly/disassembly and two articulated-object tasks. Dexora’s mean is 89.6%, versus GR00T 82.1%, π₀ 50.4% and DP 34.2%; seven of twelve tasks reach at least 90%. It leads the shaded bimanual tasks, but GR00T is better on ring stacking (90 versus 85) and ties bowl placement and square-block retrieval. Thus the abstract’s 90% is a rounded aggregate, not a universal task success rate. The basic evaluation is called zero-shot after corpus training; the paper does not detail overlap between test tasks and the 200-task training corpus.

Dexterity. With 100 task demonstrations, the six-task average is 66.7%, versus GR00T 51.7%, π₀ 26.7% and DP 6.7%. Dexora improves five tasks over GR00T, but cuts leek at 80 versus 85. Cap twisting is only 25%, though all baselines score zero. The authors identify absent tactile feedback and low-friction rigid fingertips as causes of slipping. Their 12-DoF versus baseline 6-DoF hand comparison couples model/data changes with physical dexterity, so it does not isolate the learning algorithm’s contribution.

Complete original PDF table S4.T1
TABLE I: Basic tasks evaluation. Results are success rates (%) over 20 trials. Gray columns indicate bimanual tasks.
Original selectable table
Method Pick and Place Assemble / Disassemble Articulated Object Avg.
Apple →\to plate Bowl →\to bowl Two eggs →\to box Lift basket Left block →\to right plate Stack ring blocks Grab square blocks Place kettle on base Remove pen cap Separate nested bowls Open cabinet door Open laptop
DP 60 65 30 10 25 35 15 45 30 10 65 20 34.2
π0\pi_{0} 75 70 45 30 30 60 60 65 55 20 60 35 50.4
GR00T N1 95 100 75 60 80 90 80 90 80 60 95 80 82.1
Dexora 100 100 85 80 90 85 80 95 90 80 100 90 89.6
Complete original PDF table S4.T2
TABLE II: Dexterous manipulation tasks evaluation.
Original selectable table
Method Use pen Fetch book Cut leek Place plates Rough dough Twist cap
DP 5 10 10 0 15 0
π0\pi_{0} 20 45 60 20 15 0
GR00T N1 45 60 85 60 60 0
Dexora 65 80 80 70 80 25
Complete original PDF figure S4.F7
Fig. 7: Dexterous manipulation sequences. (a) Use Pen: The left hand picks up the pen (#1), hands it to the right hand (#2); the right thumb depresses the tip (#3) and writes on paper (#4). (b) Cut Leek: The right hand grasps the knife (#1), the left hand stabilizes the leek (#2); the right hand slices (#3) and returns the knife to the table (#4). (c) Rough Dough: Both hands press the rolling pin simultaneously (#1) and push forward to flatten the dough (#2). (d) Twist Cap: The left hand holds the bottle while the right thumb–index grip twists the cap (#1) and removes it (#2).

IV-C Generalization

Environmental shifts. For apple-to-plate, Figure 8 reports 100% in the base setting and unseen background, 95% under new lighting, object or height, and 90% under occlusion or clutter. Each uses twenty rollouts. This is a one-task robustness test, not evidence for all dexterous tasks or arbitrary combinations of shifts.

Other robots. Examples cover a single-arm Panda/gripper, dual-arm Cobot Magic ALOHA grippers, and a Unitree G1 arm with Inspire Hand. Unused action dimensions are padded and absent cameras masked; every transferred task receives 100 fine-tuning demonstrations. Figure 9 shows stacking, placing, pouring and a pepper handover, but no quantitative per-embodiment success table is supplied. Dimensional reduction alone does not establish a correct mapping between distinct kinematics, and these adapted examples do not demonstrate zero-shot universal control.

Complete original PDF figure S4.F8
Fig. 8: Generalization of six Out-of-Distribution (OOD) conditions. We report success rate (%) over 20 rollouts.
Complete original PDF figure S4.F9
Fig. 9: Cross-embodiment generalization. The Dexora policy transfers to (a) single-arm gripper, (b) dual-arm grippers, and (c) single-arm single-hand, completing representative tasks like a three-step pepper handover.

IV-D Ablation Study

Data composition. Figure 10 compares simulation only, simulation plus half the real tasks, and simulation plus all real tasks. Apple-to-plate rises 75→90→100%; ring stacking 65→80→85%; pen use 0→35→65%; leek cutting 10→60→80%. The prose instead gives the last value as 85%, while the plotted label and main table say 80%. Adding real data also expands task coverage, including dexterity absent from the synthetic task set, so the experiment does not isolate realism from diversity.

Quality weighting. Table III raises corn-to-plate success from 85 to 95% and basket lifting from 55 to 80%. Normalized acceleration falls 0.034→0.020 and 0.041→0.023; jerk falls 0.043→0.032 and 0.052→0.036. Figure 11 shows selected joint traces and failures. This supports the complete weighting recipe on two tasks, without separating smoothness filtering, replay validation, compatibility features and learned scoring, or establishing reliability across seeds.

Complete original PDF figure S4.F10
Fig. 10: Effect of training data composition. Success rate for four tasks under three training regimes: Sim Only, Sim + 50% Real, Sim + All Real.
Complete original PDF figure S4.F11
Fig. 11: Effect of the data-quality discriminator. (a) Corn →\rightarrow plate: with the discriminator, joint trajectories are smooth and the placement succeeds; without it, high-frequency oscillations in left-hand joint 5 cause the corn to drop. (b) Lift basket (bimanual): with the discriminator, the basket is lifted; without it, jitter in right-hand joint 9 tilts the basket and it slips.
Complete original PDF table S4.T3
TABLE III: Effect of the discriminator model. We report S.R. (success rate %) and smoothness metrics—mean normalized joint Acceleration and Jerk, averaged over 20 episodes.
Original selectable table
Method Corn →\to plate Lift basket
S.R. Acc. ↓\downarrow Jerk ↓\downarrow S.R. Acc. ↓\downarrow Jerk ↓\downarrow
w/o discriminator 85 0.034 0.043 55 0.041 0.052
w/ discriminator 95 0.020 0.032 80 0.023 0.036

V CONCLUSION

The contribution is an integrated high-dimensional demonstration and learning system: embodiment-matched simulation provides scale, real demonstrations supply contact-rich skills, and offline scoring helps prioritize useful behavior. The observed gains come with restricted evaluation coverage, incomplete training-detail specification and important tactile-control failures. Proposed future work adds tactile feedback and longer-horizon memory, subgoals and planning; those capabilities are not demonstrated here.

Concept demonstration

Open image ↗