Latest / Embodied AI 101 / Dexora: Open-Source VLA for High-DoF Bimanual Dexterity

Transcript

Dexora: Open-Source VLA for High-DoF Bimanual Dexterity · Embodied AI 101

• Dexora: Open-Source VLA for High-DoF Bimanual Dexterity. Modern robotics has made impressive strides, but one frontier still looms large: bringing together fine manipulation and bimanual coordination in a single, end-to-end learning system. Until recently, robotic vision-language-action (VLA) models – systems that take in visual observations and language commands and output full action sequences – were typically limited to either simple dual-arm gripper control or single‐arm dexterous hands. In other words, robots could either be “two-armed but coarse” or “one-armed but very fine,” but not both at once. Dexora, a newly unveiled ICRA 2026 system, shatters this compromise by natively targeting dual-arm, dual-hand, 36-DoF manipulation. By unifying gross arm movements and nuanced finger controls, Dexora opens the door to tasks like unscrewing a bottle cap while one arm grips it and the other twists – precisely the sort of coordinated, high-DoF work that prior models could only dream of. Behind Dexora lies a massive effort in data and architecture. The core is a diffusion‐Transformer policy network trained on a hybrid dataset amassed via a custom teleoperation setup. A human operator wearing a lightweight exoskeleton for arm motion and an Apple Vision Pro headset for finger tracking drives a real dual-arm robot and a matched MuJoCo simulator in lockstep. This “dual simulator” approach yields both high-fidelity real demonstrations (over 12,000 episodes, ~40 hours) and a scale of simulated trials impossible to gather in hardware (100,000 episodes, ~6.5 million frames). Crucially, Dexora’s training pipeline is data-quality‐aware: it trains an offline discriminator to score each demo clip and then weights the diffusion-policy learning by those scores. The upshot is a robust VLA model that, in experiments, far outperforms prior baselines on both basic manipulation tasks (achieving ~90% success on simple pick-place tasks) and truly dexterous benchmarks (66.7% success where the next-best model musters only ~51.7%). Astonishingly, the 36-DoF Dexora model even generalizes to unseen robots: by padding or masking action dimensions, the same network can drive single-arm grippers, dual-claw setups, or single-arm dexterous hands without retraining. In this article, we unpack Dexora’s story in depth. We’ll explain why bimanual dexterity is a grand challenge, how Dexora’s teleoperation-and-simulation pipeline works, what its diffusion-based policy looks like, and how its data-filtering discriminator gives it an edge. Along the way, we’ll compare to related VLA and policy-learning work, and point out the avenues Dexora opens for future embodied AI. The Challenge: Bimanual Dexterous Manipulation. Humans can effortlessly coordinate two hands and many fingers to do things like unscrew a jar, assemble parts, or knead dough – tasks that require precise positioning, forceful coordination, and often complex finger movements. Replicating such behavior in robots has historically been split into two camps:. Dual-arm, gripper-based systems: Here each arm has a simple parallel jaw gripper (2–3 DOF), and tasks focus on cooperative motions like passing objects between hands. These tasks can often be solved with classical planning or simpler controllers. Indeed, many early “VLA” systems (vision-and-language conditioned controllers) concentrated on these problems and achieved success by, for example, planning pick-and-place actions for each gripper. Single-arm, dexterous-hand systems: Advances in robot hands (e.g. 12+ joint hands) allow very fine-grained manipulation with complex finger articulation. Recent work has focused on, say, a single robot arm doing intricate tasks like assembling small parts or manipulating deformable objects. These tasks clearly benefit from end-to-end learning on all modalities (vision, language, proprioception). However, these methods typically assume only one active arm and cannot naturally extend to two-handed coordination. What was missing until now is a unified VLA system that natively handles both. Think of a common household scenario: opening a jar of jam. You’d want one arm to stabilize the jar (and maybe rotate it if needed) while the other unscrews the lid, with your fingers feeling the resistance. Low-DOF grippers can’t open small lids; single-arm robots can’t hold the jar steady. Yet tasks like this are classic real-world needs. Prior to Dexora, one could cobble something together by using two independent VLA models (one for each arm) or by heuristic coordination, but robot learning stories always had this gap: systems either “could do two arms but not finesse” or “could do high finesse but only with one arm.” Dexora breaks that barrier. It is built from the ground up to handle 36 joint angles at once (two 6-DoF arms plus two 12-DoF hands, each finger fully actuated). And it returns control in one shot: given camera images and a language command, Dexora’s policy outputs the next chunk of all 36DoF actions simultaneously. The payoff is a skill set that spans a much richer set of tasks. Dexora can assemble parts requiring both arms, but also perform classic “dexterous” tasks like twisting a cap or cutting with a knife (holding something in one hand, manipulating with the other). In benchmarks, tasks that were nearly impossible for earlier single-arm systems (or purely gripper-based ones) become doable. In short, Dexora shows that “dual-armed and dexterous” is no longer an either/or in VLA. Teleoperation and Data Collection: Bridging Reality and Simulation. Training a model for 36-DoF control is a data-intensive endeavor. Dexora tackles this challenge head-on by combining human teleoperation with both real and simulated environments in lockstep. The core idea is elegantly simple: stream operator commands to both a physical robot and its simulated twin simultaneously, so that every demonstration yields two parallel datasets. Figure illustrates this setup (image from Dexora’s project page). A human operator wears a lightweight exoskeleton backpack on the torso to drive the robot’s arms (shoulder, elbow, wrist joints) and an Apple Vision Pro headset to capture finger movements. The exoskeleton measurement yields a drift-free, low-latency reading of each shoulder/elbow/wrist angle; the Vision Pro’s inside-out tracking produces a 3D skeleton of each hand in real time. Together, these map to the robot’s dual arms and two 12-joint hands, so that the operator effectively tele-operates a full 36-DoF avatar robot. All sensor streams from the real and simulated robots are recorded (for instance, four synchronized RGB cameras – one head-mounted, two wrist-mounted, one static third-person – plus the full 36 joint-state vector) at 20 Hz. In effect, every wrist motion, every finger curl, and every camera view is logged twice: once from the physical setup, and once from the exact MuJoCo replica. This ensures the resulting dataset has:. High-fidelity real demonstrations: The team collected about 10–12K episodes of teleoperated tasks, totaling roughly 2.9 million frames or about 40 hours of real robot play. The data cover a variety of scenarios, including cluttered scenes, obstruction, and challenging contacts. Using an exoskeleton (which has essentially no drift and low latency) and vision-based finger capture means the operator can move and articulate the robot very precisely. In practice, this system can produce human-quality demonstrations of even the most dexterous tasks. Large-scale simulated demonstrations: Simulations allow for enormous scale at low cost. Dexora uses MuJoCo to generate 100,000 simulated episodes (approximately 6.5 million frames, or 361 hours). Crucially, these are embodiment-matched to the real platform: the physical robot’s kinematic and dynamic properties are replicated in software. Because every teleop command is also sent to the MuJoCo twin, the sim data are not mere hand-crafted training runs but strict analogues of the real demos. After collecting a few seed demonstrations (3–5 human drive examples per task), Dexora applies a “DexMimicGen” procedure: randomize object initial states and retarget those seed actions in simulation to generate hundreds more variations per task. The result is a complementary dual dataset. The simulation data provide breadth (100K episodes with 361h runtime across 30 object categories and 200 task variants), while the real data provide ground-truth reality and natural noise. By coupling physical and virtual collection, Dexora mitigates the reality gap: the same tasks appear in sim and real under identical high-DOF kinematics and multi-view sightlines. It’s a virtuous cycle – the real data ensure realism (e.g. real object textures, unexpected collisions) while the sim data ensure volume and variety. Overall Dexora’s corpus includes something like 300 manipulated objects (297 in sim, 347 real) across dozens of categories, covering both simple work [“pick up a block”] and complex handcraft [“unscrew a cap”]. In sum, the teleoperation pipeline is Dexora’s secret sauce for data. By designing an integrated exoskeleton+VR setup and logging everything on a matched digital twin, the team produced orders of magnitude more training data than would be feasible by physical teleop alone. This massive dataset is freely released (“libero-2.1” format) to the community, marking Dexora as the first public VLA dataset at this scale for high-DOF bimanual manipulation. Model Architecture: Diffusion-Transformer Policy. With the data in hand, Dexora needs a policy powerful enough to map from raw vision and language inputs to 36 continuous motor actions over time. The team chose a diffusion-based Transformer architecture, building on recent successes of diffusion policies in robotics. Conceptually, the network is a decoder-only Transformer that denoises action sequences conditioned on vision, language, and proprioceptive state stacks. During inference, Dexora’s policy operates as follows. Every few timesteps, given the robot’s current state $s_t$ (the joint angles/velocities for all 36 dimensions), the latest multi-view RGB observations $o_t$ (concatenating, for example, the egocentric head camera plus the two wrist cameras) and the text instruction $\ell$ (e.g. “open the red box” or “twist the blue cap clockwise”), the model outputs a sequence of future actions for a fixed horizon. Practically, it predicts the next $L$ actions in a single shot via the diffusion process: starting from random noise, the model iteratively “denoises” to fill in a smooth action trajectory that accomplishes the command. This differs from standard auto-regressive policies by generating a whole chunk of actions simultaneously, which is helpful for coherence in high-DoF spaces. On the input side, Dexora uses modern encoders for each modality: a pre-trained T5-like text encoder handles the language instructions, and a visual encoder (the project refers to a “SigLip” encoder, presumably a ViT-like model) turns images into token embeddings. These embeddings are then woven together with the current joint-state vector in a stack of Transformer blocks. Inside these blocks, “cross attention” tags between vision, text, and state allow the model to align, say, the word “cap” with the cap object in view and compute how arms and finger joints must move jointly. The output layers of the Transformer produce estimates of the Gaussian diffusion noise at each of the next $L$ time steps. In practice, Dexora’s action space is 36 continuous dimensions, so each predicted action is a 36-vector of joint angle velocities (or similar) for the robot. Key to Dexora is end-to-end learning of this entire vision-to-action mapping. Unlike classical pipelines that might break the problem into perception and planning submodules, Dexora’s diffusion model sees pixels and language and churns out joint torques all within one learned network. This fully differentiable fusion helps it exploit subtle cues (for example, if the word “twist” appears, the model can focus on the twisting degree of freedom in the wrist joints). Importantly, the policy must learn sequential reasoning too: multi-step tasks (e.g. “pick up the block, then insert it” or “grasp bottle, then rotate cap”) require the diffusion steps to generate a temporally coherent plan. The model’s large-scale pretraining (to describe below) equips it with over 400 hours of robotic motion data, giving it a strong prior for smooth sequences. In sum, Dexora’s architecture is a cutting-edge multimodal policy network: a diffusion Transformer conditioned on text and images, outputting fine-grained 36-DoF motions. Similar diffusion policies have appeared recently (e.g. “Decision Diffuser” or “Diffusion Policy” work), but Dexora applies the idea at a scale and embodiment never done before. Integrating both arms and both hands seamlessly is what distinguishes its policy network from prior single-arm diffusion policies. Training Recipe: Quality-Aware Imitation Learning. Training a diffusion policy on imperfect human data is tricky. Human teleoperation is precise, but inevitably noisy: hands slip, the operator hesitates, or cameras jitter. On rich real robot data, these imperfections can pull the model off the desired solution. To tame this, Dexora introduces a clever data-quality discriminator into the training loop. The goal is to learn which parts of the human demonstrations are “good” and which should be down-weighted during learning. The training is broken into three main phases, each feeding into the next:. Simulation Pretraining: First, Dexora trains the diffusion Transformer exclusively on the 100K simulated trajectories. These simulations are clean and diverse (objects from Objaverse with varied physics parameters) and cover all the basic behaviors the policy should know (grasping, placing, rotating, assembling simple mechanisms). This stage boosts the policy’s competence for basic tasks. The policy learns a general mapping from vision and instruction to gross actions, without yet worrying about real-world noise. Even here, the multi-DoF nature of the data is a challenge, but with so much sim data (far more than one could practically collect by teleop alone), the network can solidify fundamentals like “how to reach to an object” or “how to lift and retract without collisions.”. Discriminator Training (Quality Modeling): Next, attention turns to the real-world data, which is richer but messier. Dexora takes a subset of the real demonstrations and computes hand-crafted quality metrics for each episode. The team uses measures like the smoothness of the motion (low acceleration and jerk) and simple success checks (did the open-loop replay complete the task without collisions?). They select the top ~20% smoothest episodes as candidate “high-quality” data. Then they further verify those by replaying them on the robot to ensure they truly succeed, preserving the best ~15% in a high-quality set $S_{high}$. With that subset, Dexora trains a small quality discriminator network. The input to this discriminator is a short clip of demonstration (multi-view frames + states + the chunk of actions + even a proxy “log probability” from the current diffusion policy). The network itself is a shallow Transformer that outputs a score $d(C)\in(0,1]$ for each clip. During training, clips from $S_{high}$ are treated as “positive” examples and others as “unlabeled” (you don’t know if they’re good or bad). Using a positive-unlabeled loss, the discriminator learns to give high scores to clips that look like the smooth, successful ones and low scores to the noisier clips. After training, this discriminator basically predicts “how good” any real-data clip is, according to the model’s learned standards of motion and task completion. Data-Weighted Fine-Tuning: Finally, the main diffusion policy is fine-tuned on the entire real dataset (both the high-quality and less-perfect demos), but with an important twist: each clip is weighted by the discriminator’s score. In practice, you compute the discriminator score $d(C)$ for each clip and clamp it to, say, [0.1,0.9], then convert that to a weight $w_i$ for the imitation loss. Clips deemed high-quality by the discriminator have $w_i$ near 1, while shakier clips get very low weight. During policy training, the standard diffusion loss $|\epsilon_\theta - \epsilon|^2$ (the usual denoising MSE) is multiplied by $w_i$ for each clip. This means the policy “pays attention” mostly to the good data and essentially ignores the worst, thus avoiding being dragged toward poor examples. Crucially, at inference only the policy is used (the discriminator is offline), so the result is a cleaner learning signal without online changes to the policy’s behavior. Collectively, this three-phase recipe yields a robust model. Pretraining on sim gives breadth and a learning foundation; the discriminator ensures that flawed real demos don’t corrupt training; and the weighted fine-tuning lets the model specialize on real nuances (like actual friction or object appearances) where it counts. Dexora’s authors did thorough ablations: both simulation pretraining and the weighted approach turn out to be essential for high performance. Training on raw real data without quality weighting leads to choppy, jittery policies that stumble on fine tasks, whereas using the filtered post-training regime yields smooth, stable hand motions. In essence, Dexora’s training is “quality-$A_{ep}+J_{ep}$ aware” – it knows when to trust a demonstration and when to ignore it. This approach follows a growing trend in offline imitation learning. Prior works (like Diffusion Policy by Woo et al. 2023) have shown that big data policies benefit from denoising and weighting strategies. Dexora adapts that idea to the specifically noisy realm of teleoperation data, using the hybrid pipelines and Transformer power to make sense of the mess. The offline nature also sidesteps safety issues – the policy never learns via trial-and-error on the actual robot, only via these recorded episodes. Experimental Results: Task Success and Robustness. How well does Dexora actually perform? The ICRA paper reports a broad evaluation suite of tasks grouped into basic vs. dexterous categories, as well as tests for generalization. The bottom line is clear: Dexora outclasses past approaches across the board, especially on tasks that truly demand its dual-arm dexterity. Basic manipulation tasks: These include things like pick-and-place with one or both arms, simple assembly/disassembly, and articulated object manipulation (e.g. opening a drawer or turning a rigid handle). In these tasks, Dexora achieves an average success rate of about 90%. In the published table, 7 out of 12 basic tasks exceed 90% success. Notably, bimanual coordination tasks – for example “hand off a tool from left hand to right hand” or “un-nest a pair of bowls” – show Dexora’s advantage. These tasks are sometimes actively avoided by simpler policies, but Dexora solves them reliably. For instance, in a experiment drawing from the Chinese summary: “双臂协同任务（如双手递物、嵌套碗分离）.[Dexora] far exceeds GR00T N1 (82.1%) and π0 (50.4%) baselines”. That is, on two-arm coordinated tasks (like passing objects or separating nested bowls), Dexora is markedly more successful than a top-of-line baseline (only about 82% for GR00T vs Dexora’s ~90%). On non-coordinated tasks (single-arm reaches, simple picks), Dexora is already near ceiling. Dexterous fine manipulation tasks: Here we see Dexora’s punch. These tasks involve intricate finger movements or twists: examples include Twist Cap (open a bottle lid), Use Pen (possibly unscrew or move a pen like a tool), Rough Dough (manipulate a deformable mass), Cut Leek and Cut Cucumber (one hand holds the knife, other the food), etc. On these, Dexora’s average success is 66.7%, whereas the best competing model only managed 51.7%. In particular, tasks that require both arms and finger finesse see the biggest gaps. The review article notes that in tasks like unscrewing a cap or shaping dough – which “need two arms plus multi-finger control” – other baselines “almost fail” but Dexora still completes them stably. The baselines they compare to include:. A “Diffusion Policy (DP)” model, presumably a single-arm diffusion controller applied to each arm. “$\pi_0$” and “GR00T N1”, which refer to generalist robot models from prior work. GR00T is an NVIDIA foundation model (an “Isaac-GR00T N1.7” style policy) and $\pi_0$ appears to be another recent VLA-like model. In tough tasks, those models either don’t support the needed degrees of freedom or weren’t trained on such interplay. By contrast, Dexora’s dual-arm, dual-hand design means it inherently has the capability to attempt these moves. Cross-embodiment generalization: Perhaps the most surprising result is that Dexora can be deployed on robots different from its training platform. Because the policy outputs a generic 36-DoF action vector, one can adapt it to a lower-DoF robot by ignoring (masking) the unused dimensions. For example, a single-arm gripper robot can run the Dexora model by zeroing out the other arm and hand commands. The team measured that even in such cases, the model works “out of the box” without any extra training. This is because the controller’s knowledge of how to move the arms and hands isn’t hardwired to specific limbs – it’s encoded in language/vision context and action semantics. The result is a “34-DoF” or “6-DoF” robot that can still use the 36-DoF policy by simple dimension-remapping. This is a big advantage of Dexora’s unified approach: you get One Policy to rule them all. In experiments they showed Dexora running on:. Single-arm with gripper (other DOFs masked off). Dual-arm each with only a 2-finger claw. Single-arm with one XHAND (and rest zero). In each case, no retraining was needed, and performance dropped only slightly. This “downward compatibility” breaks a usual robotics rule – that models must match the robot exactly. Dexora suggests that training big and then shipping small is possible for VLA. Out-of-distribution robustness: The team also tested Dexora under visual changes it never saw during training: novel lighting, backgrounds, unseen distractor objects, occlusions (e.g. part of the scene blocked), and even while moving the camera height. Dexora’s success rates dipped only marginally in these scenarios. In one figure it maintains high success across six such OOD conditions. This robustness likely stems from its large, varied training set and its end-to-end conditioning on vision-language (which has some built-in invariances through the visual encoder). In practice, Dexora isn’t fooled by, say, a new wallpaper or a partly-hidden tool – it still does the right motion with high probability. Ablation studies: To verify their design choices, the authors include “ablation” experiments. Not surprisingly, training without the data-quality discriminator leads to jittery, unreliable hand motions. Policies collapse more often or generate implausible finger wiggles. Likewise, training without any real data (only sim) yields a model that struggles on tactile or complex contacts. The healed Dexora – with the full sim+real and discriminator – is markedly superior in dexterous tasks. These analyses underline the conclusion that real demonstrations + quality weighting are the core of its dexterity gains. In summary, in extensive tests Dexora consistently outperforms previous VLA baselines. It achieves near-perfect success on basic tasks, 15 points higher success on dexterous tasks than the best competitor, and maintains generalization to new robots and conditions. These empirical results back up the bold claim: unifying bimanual and dexterous control in one model really does pay dividends. Related Work and Positioning. Dexora arrives at a moment when the robotics research community is converging on “foundation model” paradigms. Just as large language models showed that scaling data and network size can produce broad linguistic abilities, roboticists are seeking analogous giant multi-task policies. Dexora cites and builds on many of these trends:. Diffusion-based policy learning: The idea of using diffusion models (denoising probabilistic models) to generate robot actions was explored recently by works like “Diffusion Policy” (Choi et al., 2022/3) and “InstructBERT for Actions”. Dexora’s policy network is a direct descendant, where a Transformer “denoises” action sequences given context. Compared to earlier diffusion policies, Dexora just throws in far more degrees of freedom and tasks. Vision-Language-Action (VLA) systems: Dexora positions itself as the next step beyond first-gen VLA: prior VLA works (for example, CLIPort, VIMA, and various vision-language grasping projects) mostly handled at most one arm or simple two-arm with grippers. For instance, Meta’s VIMA (Ravens dataset and follow-ups) allowed language-conditioned tasks on a single 7-DoF arm with a parallel gripper. Dexora is the first to label itself “VLA” with support for two arms and dexterous hands. The Dexora authors explicitly cite DexVLA (Wen et al., 2025) as a related idea – DexVLA is also a diffusion-based VLA, but focused on generalist humanoid tasks. Dexora is essentially the “dual-handed extension” of that trend, and the first open source of its kind. Generalist robot policies: Systems like NVIDIA’s Isaac-GR00T (Ge et al.) and the Psi-Zero project (Physical-Intelligence Lab) are working on giant policies for many robots. Dexora compares itself to the GR00T baseline (calling it GR00T-N1 in results). GR00T is a GPLM (general-purpose language model) for robots that can output actions in different embodiments, but to our knowledge it has not been focused on 12-DoF hands. Dexora’s empirical gap against GR00T suggests that specialized high-DoF training brings clear benefits. Similarly, the Psi-Zero line (Ψ₀) is about a humanoid-level VLA but again those efforts are either in simulation or limited hardware, whereas Dexora is focused on a physical platform. Offline RL and Imitation Learning: Beyond diffusion, Dexora ties into a large body of work on training on logged data. Its three-stage recipe (sim pretraining, filtering, fine-tuning) echoes best practices from offline RL (pretrain on a diverse corpus, filter by dynamics consistency, finetune). In fact, Dexora’s discriminator-weighting is analogous to some offline RL techniques that down-weight outliers or unreliable data. But Dexora remains an imitation approach – it does not use a reward function – which aligns it more with modern “imitation with noisy expert data” research, rather than full RL. Datasets for dexterous control: Another angle is that Dexora itself contributes what might become the largest open dataset of bimanual dexterous manipulation. Previous datasets (e.g. Adroit, ManiSkill, etc.) have dexterous hands but not dual arms. The Dexora data collection, especially with its synchronized simulation records, could spur new research. It joins other large-scale robotics datasets (like the LIBERO dataset, which Dexora mentions) as benchmarks for training multimodal policies. In summary, Dexora is both a continuation and a leap in these trends. It continues the vision-language policy paradigm of recent years, but leaps by scaling up the embodiment. It leverages diffusion and offline policy ideas, yet its focus on dual-arm dexterity is novel. Early works hinted that maybe one could get a single network to do everything, but Dexora is the first to actually demonstrate that unification on real hardware across dozens of tasks. It is encouraging to see the patterns: diffusion networks, transformer architectures, and large human-collected corpora are becoming standard in robot control. Broader Impact and Future Directions. Dexora is a tour de force, but it has limitations and it naturally suggests future work. The authors themselves note a few key caveats:. Hardware complexity: A 36-DoF robot with exoskeleton teleop is not cheap or easily replicated. Dexora’s setup cost (two AIRBOT arms and two SH-XHAND hands, plus the custom exoskeleton rig) is substantial. This may slow immediate adoption by other labs, limiting reproducibility to well-funded teams. However, by open-sourcing the design and data, Dexora mitigates this: other groups could at least simulate an identical setup or mount off-the-shelf hardware. Lack of touch sensing: Real human hand skills rely heavily on tactile feedback (feeling the twist of a cap, sensing friction). Dexora’s system has no tactile channel – everything is done on visual cues and joint positions. The Chinese report notes that tasks like unscrewing a bottle cap still have a noticeable failure rate, likely due to missing force feedback. Integrating touch sensors on the fingertips, or force feedback in the exoskeleton, could be a major improvement for the future. Long-horizon tasks: Dexora’s episodes are relatively short, one-shot tasks. It does not (yet) tackle long, multi-step procedures like assembling a series of components from a toolbox. Those would require memory, planning over tens of seconds, and perhaps external knowledge. Extending the VLA framework with a persistent state or a higher-level planner is an open challenge. Extreme generalization: While Dexora handles moderate distribution shifts, the real world has no end. Situations like totally novel materials (liquids, very slippery objects), or rapid dynamic change (moving objects), may still break it. Also, it was trained on relatively clean teleop; more chaotic human scenes (with more clutter or other humans) might confuse it. So broadening the dataset with even more real-world diversity would help. Compute and scalability: Training diffusion transformers on multi-million-frame data requires substantial compute. Not every lab can replicate that. That means the barrier to entry is high. However, Dexora’s open model weights (if released) or training code could help others fine-tune rather than train from scratch. Looking forward, Dexora’s impact will be two-fold. On the research side, it sets a new benchmark: any future dexterous VLA model will be compared against it. We can already imagine follow-ups: perhaps a DexoraV2 with more arms, or a similar idea for quadruped walking+manipulation, or combining it with LLMs for on-the-fly plan generation. Roboticists might plug Dexora’s policy into new scenarios (e.g. service robots in homes). In applications, it hints at what might be possible: imagine assembly-line tasks where two robot hands perform intricate operations without human teleoperators on site, guided only by language instructions. For our field, Dexora underscores a general lesson: the combination of large-scale data, hybrid sim-real pipelines, and powerful sequence models can engulf prior limitations. Tasks once considered out of reach (bimanual human-like dexterity) can be learned by scaling up. It’s reminiscent of the successes in computer vision and NLP. We should nevertheless remember that Dexora is a model learned from data, not a guarantee of perfect success. Any deployment in safety-critical contexts would still need oversight. But as a demonstration, Dexora is a leap toward the “universal manipulator.”. In the context of embodied AI, Dexora is a hallmark achievement of 2026. It shows open-source can lead in what might have been expected to be proprietary (NVIDIA, OpenAI, etc.). By releasing code and data, the authors accelerate further progress. Over this decade, I anticipate many projects will cite Dexora as the first to crack the bimanual-dexterity nut. Perhaps in 2030 we’ll look back and say the labs that first built on Dexora’s foundation are the ones who brought truly general-purpose robotic manipulation into reality. References: Dexora’s results and methods are documented in the official ICRA 2026 paper (arXiv:2605.18722). The GitHub and project website provide a wealth of implementation detail. For context, see also Wen et al.’s DexVLA (2025) on diffusion policy in VLA systems, and NVIDIA’s Isaac-GR00T as a baseline generalist robot model. The Chinese technical reviews and English blog article summarize Dexora’s key points. All citations here refer to those public sources.