RoboTwin: Dual-Arm Manipulation Benchmark
- RoboTwin is a dual-arm benchmark and simulation platform that synthesizes digital twins from 2D images to generate diverse, realistic data for bimanual manipulation.
- It employs spatial relation-aware code generation and structured domain randomization, improving manipulation success rates by up to 10.9 percentage points.
- RoboTwin enhances sim-to-real transfer, with experiments showing up to a 367% relative improvement on unseen, cluttered tasks.
Searching arXiv for recent RoboTwin papers to ground the article. {"query": "RoboTwin benchmark dual-arm manipulation arXiv", "max_results": 10} {"query": "\"RoboTwin 2.0\" arXiv", "max_results": 10} {"query": "\"RoboTwin\" teleoperation digital twins arXiv", "max_results": 5} RoboTwin is presented in the manipulation literature as a dual-arm robot benchmark and simulation platform designed specifically for generalizable bimanual manipulation, and in its later 2.0 form as a scalable simulation framework that enables automated, large-scale generation of diverse and realistic data together with unified evaluation protocols for dual-arm manipulation (Chen et al., 29 Jun 2025). The benchmark lineage combines generative digital twins, spatially grounded task synthesis, structured domain randomization, and matched simulation–real evaluation for dual-arm systems (Mu et al., 17 Apr 2025, Chen et al., 22 Jun 2025). The name is not entirely univocal, however: a separate paper uses “RoboTwin” for a dual-digital-twin telesurgery teleoperation framework rather than a manipulation benchmark (Yelchuri et al., 1 Jun 2025).
1. Origins, scope, and naming
The benchmark lineage begins with the early RoboTwin system, which introduced a generative digital twin framework that uses 3D generative foundation models and LLMs to produce diverse expert datasets and provide a real-world-aligned evaluation platform for dual-arm robotic tasks (Mu et al., 2024). In that formulation, RoboTwin creates varied digital twins of objects from single 2D images, introduces a spatial relation-aware code generation framework, and validates the resulting pipeline on the open-source COBOT Magic Robot platform (Mu et al., 17 Apr 2025).
RoboTwin 2.0 extends that agenda from a benchmark with generative digital twins into a scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation (Chen et al., 22 Jun 2025). It adds RoboTwin-OD, automated expert synthesis with multimodal LLMs and simulation-in-the-loop refinement, embodiment-aware grasp adaptation, and a substantially larger task and embodiment coverage. Later work then treats RoboTwin 2.0 as a primary evaluation environment for VLA models, world-action models, distillation methods, augmentation pipelines, and robustness audits (Cai et al., 8 Jun 2026, Fan et al., 13 Apr 2026, Jiang et al., 2 Jun 2026).
A distinct usage of the same name appears in a surgical teleoperation paper, where RoboTwin denotes a dual digital twin framework for remote robotic surgery; there the emphasis is minimum perceived latency, safety mediation, and reduced network traffic rather than manipulation benchmarking (Yelchuri et al., 1 Jun 2025). In current robotics usage, context therefore matters.
| Usage | Defining characteristics | Representative source |
|---|---|---|
| Early RoboTwin benchmark | Generative digital twins from single 2D images; spatial relation-aware code generation; matched sim and real evaluation | (Mu et al., 17 Apr 2025) |
| RoboTwin 2.0 | RoboTwin-OD, automated expert synthesis, five-axis randomization, 50 tasks, 5 embodiments, 100k+ trajectories | (Chen et al., 22 Jun 2025) |
| CVPR 2025 challenge ecosystem | Three-stage competition built on RoboTwin 1.0/2.0 and AgileX COBOT-Magic | (Chen et al., 29 Jun 2025) |
| Telesurgery RoboTwin | Dual digital twins for low-latency teleoperation and safety | (Yelchuri et al., 1 Jun 2025) |
2. Generative digital twins and automated expert synthesis
A central RoboTwin idea is that object diversity should be synthesized rather than assumed. In the early benchmark, a single RGB image of a real object is processed by GPT-4V for textual description, diversified through additional language generation, rendered into multiple 2D variants with Stable Diffusion XL Turbo, and then converted into 3D assets by the image-conditioned 3D generator Rodin; the resulting assets include geometry, surface normals, and textures and are imported into SAPIEN/ManiSkill3 as physics-compatible objects (Mu et al., 17 Apr 2025). RoboTwin 2.0 systematizes this into RoboTwin-OD, a large-scale object library comprising 731 instances across 147 categories, including 534 in-house reconstructed objects, 153 from Objaverse, and 44 articulated objects from SAPIEN PartNet-Mobility, with manipulation-relevant annotations and 15 natural descriptions per object (Chen et al., 22 Jun 2025).
The benchmark is not only an asset library; it is also an expert-data factory. RoboTwin 2.0 uses a task code generation module in which multimodal LLMs produce Python task programs over a manipulation API, while a simulation executor and a VLM observer diagnose failures and feed structured corrections back into the code-generation loop (Chen et al., 22 Jun 2025). Each generated program is executed 10 times; the loop terminates when success rate exceeds 50% over those 10 runs in one iteration or after 5 iterations (Chen et al., 22 Jun 2025). On a 10-task evaluation, the RoboTwin 2.0 pipeline with multimodal feedback reports 71.3% ASR, compared with 62.1% for RoboTwin 2.0 vanilla generation, and yields a 10.9 percentage-point ASR gain over the corresponding RoboTwin 1.0 multimodal-feedback setting (Chen et al., 22 Jun 2025).
Spatial structure is explicit throughout the pipeline. The early RoboTwin work defines functional points, contact points, a function axis, an approach axis, and a lateral axis for each object or tool, and uses these annotations to generate end-effector goals and task code (Mu et al., 2024). Annotation transfer across generated instances is automated by diffusion-feature correspondence: an annotated source pixel is mapped to a target pixel by maximizing Stable Diffusion feature similarity, and the matched point is then projected back to the target mesh (Mu et al., 17 Apr 2025). This spatial annotation layer is what makes the later “spatial relation-aware code generation framework” operational rather than purely linguistic.
Embodiment awareness enters at grasp generation time. RoboTwin 2.0 precomputes multiple grasp candidates with differing approach axes, perturbs them toward higher-reachability directions for a given robot, and filters them with Curobo motion planning, which materially benefits low-DoF platforms: across 50 tasks and five arms, average planning success rises from 52.2% in RoboTwin 1.0 to 60.5% in RoboTwin 2.0, with Piper improving from 2.4% to 25.1% and Aloha-AgileX from 65.1% to 78.8% (Chen et al., 22 Jun 2025).
3. Benchmark structure, data regimes, and evaluation protocols
RoboTwin 2.0 instantiates its framework across 50 dual-arm tasks spanning five robot embodiments: Franka, UR5, Piper, ARX-X5, and Aloha-AgileX (Chen et al., 22 Jun 2025). The task suite covers grasping and pick-and-place, stacking, articulated-object operation, handover, tool use, and dynamic actions such as bottle shaking; named tasks include Pick Dual Bottles, Place Dual Shoes, Open Laptop, Open Microwave, Turn Switch, Press Stapler, Stack Blocks Three, Stack Bowls Three, Handover Block, and Handover Mic (Chen et al., 22 Jun 2025, Cai et al., 8 Jun 2026).
The defining environmental mechanism is five-axis domain randomization: clutter, lighting, background textures, tabletop height, and language instructions (Chen et al., 22 Jun 2025). Clutter is sampled from RoboTwin-OD with collision-aware placement, textures are drawn from a 12,000-texture library derived from 1,000 prompts and 20,000 generated images, lighting varies in color, type, intensity, and direction, tabletop height is randomized within a small range, and instructions are composed from 60 task templates together with 15 object descriptions per object (Chen et al., 22 Jun 2025). This is the substrate later papers refer to as the “Hard” or “Randomized” setting.
The released dataset and the later benchmark protocol are related but not identical. RoboTwin 2.0 reports 100 clean trajectories and 400 domain-randomized trajectories for each task–embodiment pair, yielding over 100,000 trajectories overall (Chen et al., 22 Jun 2025). Several later multi-task evaluations on the 50-task benchmark instead use 50 clean and 500 randomized demonstrations per task, trained jointly across all tasks and evaluated with 100 rollouts per task in both clean and randomized settings (Cai et al., 8 Jun 2026, Lin et al., 23 Feb 2026, Li et al., 15 Jun 2026). This protocol, rather than the raw trajectory count alone, became the standard setting for many comparative VLA and WAM studies.
The ecosystem also includes an explicit competition protocol. The RoboTwin Dual-Arm Collaboration Challenge at the CVPR 2025 MEIS Workshop was built on RoboTwin 1.0, RoboTwin 2.0, and the AgileX COBOT-Magic Robot platform, and comprised Simulation Round 1, Simulation Round 2, and a final Real-World Round (Chen et al., 29 Jun 2025). Participants tackled 17 dual-arm manipulation tasks spanning rigid, deformable, and tactile-based scenarios; the challenge attracted 64 global teams and over 400 participants (Chen et al., 29 Jun 2025). This competition format turned RoboTwin from a dataset generator into a public benchmarking ecosystem.
4. Methodological ecosystem built around RoboTwin
RoboTwin is less a single leaderboard than a family of protocols. Reported numbers therefore belong to distinct configurations, such as 8-task challenge settings, 19-task hard subsets, 16-task clean subsets, 12-task horizon analyses, or the 50-task clean/randomized benchmark. Because these task counts and evaluation regimes differ, the percentages below are protocol-specific rather than directly interchangeable.
| Model | RoboTwin protocol | Reported result |
|---|---|---|
| AnchorDP3 | 8-task challenge setting with extreme randomization | 98.7% average success in simulation (Zhao et al., 24 Jun 2025) |
| AIM | 50-task RoboTwin 2.0 benchmark | 94.0% Easy, 92.1% Hard, 93.1% combined average (Fan et al., 13 Apr 2026) |
| AHA-WAM | 50-task RoboTwin benchmark | 93.40% Clean, 92.20% Randomized, 92.80% average; 24.17 Hz closed-loop control (Cai et al., 8 Jun 2026) |
| ACE-Ego-0 | 50-task RoboTwin 2.0 benchmark | 91.12% Easy, 90.62% Hard (Li et al., 15 Jun 2026) |
| Pose-VLA | 50-task RoboTwin 2.0 benchmark, RGB-only evaluation | 81.30% Easy, 80.72% Hard with action expert initialization (Lin et al., 23 Feb 2026) |
| AdaMoE | 19 RoboTwin 2.0 hard-setting tasks | 40.4% 49.7% average success (Shen et al., 16 Oct 2025) |
| Flash-WAM | RoboTwin 2.0 distillation setting | 85.54% average at 1v/2a; 348 ms per chunk at 1v/1a (Akbari et al., 3 Jun 2026) |
The diversity of algorithmic questions tested on RoboTwin is as important as the headline numbers. AnchorDP3 uses simulator-supervised segmentation, task-conditioned encoders, and affordance-anchored sparse keyposes in the dual-arm challenge setting, and explicitly treats keyposes rather than dense controls as the supervision target (Zhao et al., 24 Jun 2025). AIM introduces aligned spatial value maps and intent-causal attention inside a unified world-action model, arguing that future video alone is an insufficient control interface for contact-rich RoboTwin tasks (Fan et al., 13 Apr 2026). AHA-WAM reorganizes world-action modeling around asynchronous horizons, separating a low-frequency video planner from a high-frequency action executor and using Observation-Guided Video-Context Routing to keep the action stream aligned with current observations (Cai et al., 8 Jun 2026). ACE-Ego-0 uses RoboTwin 2.0 to validate a unified action representation over camera-space actions, morphology conditioning, and time-aligned chunking, together with reliability-aware supervision from egocentric human videos (Li et al., 15 Jun 2026).
A second line of work uses RoboTwin to study data, robustness, and inference mechanics rather than core policy architecture. A video-transfer augmentation framework improves RDT-1B by 8.0 absolute points on Robotwin 2.0 multi-task Hard evaluation and by 10.0 points in single-task Hard average, while preserving task semantics and action trajectories through depth-conditioned video synthesis (Hui et al., 4 May 2026). DVAC, a test-time chunking method based on denoising variance, raises a -based policy from 0.359 to 0.416 average success on 16 RoboTwin tasks (Feng et al., 2 Jun 2026). ElasticFlow studies one-step flow policies on a 12-task RoboTwin horizon analysis and reports 70.1% overall success with approximately 71 Hz 1-NFE control (Chen et al., 9 May 2026). RoboTwin 2.0-Plus, an internal robustness suite built on RoboTwin 2.0, evaluates perturbations in camera, robot state, language, lighting, background, noise, and layout, with LingBot-VA achieving 74.2% total success and exposing a characteristic WAM trade-off: strong visual robustness but lower robustness to camera and robot-initial-state shifts (Zhang et al., 23 Mar 2026).
5. Sim-to-real role and challenge performance
From the beginning, RoboTwin was meant to bind simulation and real hardware rather than replace one with the other. The early benchmark pairs 15 tasks with 100 simulated expert trajectories and 20 real teleoperated trajectories per task on the COBOT Magic platform (Mu et al., 2024). In that setting, policies pre-trained on RoboTwin-generated data and fine-tuned with limited real-world samples improve the success rate of over 70% for single-arm tasks and over 40% for dual-arm tasks compared to models trained solely on real-world data (Mu et al., 17 Apr 2025).
RoboTwin 2.0 makes the sim-to-real claim more explicit and more difficult. In real-world experiments on a COBOT-Magic dual-arm platform, a VLA model trained with 10 clean real demonstrations and augmented with 1,000 RoboTwin 2.0 synthetic trajectories per task improves from 9.0% to 42.0% on unseen-scene cluttered evaluation, which the paper reports as a 367% relative improvement (Chen et al., 22 Jun 2025). In the same study, zero-shot models trained solely on 1,000 synthetic trajectories per task achieve a 228% relative gain on the hardest unseen real-world setting (Chen et al., 22 Jun 2025). These numbers are central to RoboTwin 2.0’s claim that domain-randomized synthetic supervision can serve not merely as pretraining regularization but as a substantial source of real-world robustness.
The public challenge confirms both the promise and the difficulty of this agenda. The CVPR 2025 RoboTwin challenge used simulation rounds for rigid-body tasks and a real-world round for pouring, folding, stacking, and pen capping; despite high simulation performance, the real-world round remained difficult, with low average scores across several tasks and a best overall team score of 26.4/100 (Chen et al., 29 Jun 2025). The challenge report identifies data quality, multimodal fusion, instruction grounding, and preprocessing as practical determinants of transfer quality, and it notes that policies often make substantial partial progress before failing at final precision steps (Chen et al., 29 Jun 2025). RoboTwin therefore functions not only as a pretraining environment but also as an instrument for exposing the last-mile difficulty of real dual-arm manipulation.
6. Benchmark validity, limitations, and open directions
A 2026 benchmark audit places RoboTwin 2.0 in a relatively favorable position among manipulation benchmarks. Under the paper’s shortcut-solvability diagnostic, a DINOv2+MLP probe without a language encoder reaches 60.4% on RoboTwin 2.0 Clean and 59.4% on Randomized, well below X-VLA at 72.8% and MotuBrain at 95.8% and 96.1%, which the authors interpret as evidence that there is no obvious small-model shortcut to SOTA on RoboTwin 2.0 (Jiang et al., 2 Jun 2026). Under the same audit’s top-line significance analysis, 73.7% of RoboTwin 2.0 SOTA claims are provably statistically significant from aggregate numbers alone, a much higher fraction than reported for LIBERO or SimplerEnv (Jiang et al., 2 Jun 2026). The same audit does not run a creeping-overfitting diagnostic on RoboTwin 2.0, so that question remains open (Jiang et al., 2 Jun 2026).
The benchmark lineage also states its own limitations. RoboTwin 2.0 reports that some tasks remain unresolved by code generation, including Open Laptop, Open Microwave, Shake Bottle, and Press Stapler, and it notes remaining limitations in physics fidelity, the restriction to five robot embodiments, and the predominance of tabletop manipulation (Chen et al., 22 Jun 2025). The challenge report argues that current endpoint-biased evaluation can undervalue near-success and suggests progress-aware scoring as a future direction (Chen et al., 29 Jun 2025). Later method papers add more specific bottlenecks: AnchorDP3 identifies perception ambiguity under clutter, binary rather than part-level segmentation, and the absence of reported real-robot results in its own study (Zhao et al., 24 Jun 2025), while the RoboTwin 2.0-Plus robustness study shows that camera and robot-state perturbations remain particularly difficult even for strong world-action models (Zhang et al., 23 Mar 2026).
These findings make RoboTwin less a closed benchmark than a structured research program. Its recurrent themes are clear: large-scale simulated expert generation, stronger affordance grounding, embodiment-aware representations, longer-horizon coordination, efficient closed-loop inference, and more faithful tests of robustness under scene and embodiment shift. Later RoboTwin-based work suggests multiple technical directions for that program—open-vocabulary affordances and real2sim2real automation in AnchorDP3 (Zhao et al., 24 Jun 2025), explicit spatial intent in AIM (Fan et al., 13 Apr 2026), asynchronous planner–executor coupling in AHA-WAM (Cai et al., 8 Jun 2026), camera-space and morphology-aware pretraining in ACE-Ego-0 (Li et al., 15 Jun 2026), and modality-aware distillation in Flash-WAM (Akbari et al., 3 Jun 2026). The common implication is that RoboTwin’s lasting significance lies not only in its datasets or scores, but in its role as a digital-twin-centered experimental substrate for robust bimanual manipulation research.
