- News
Three AIRoA Research Papers and a Demo Proposal Accepted to CoRL 2026
2026-10-28
Three AIRoA research papers and a demo proposal featuring YUBI have been accepted to the Conference on Robot Learning (CoRL) 2026.
During the conference, AIRoA will present the accepted research papers and conduct an interactive demonstration in the official CoRL Demo Area on November 9 and 10. AIRoA will also organize the workshop “The UMI Arena: From Handheld Devices to Real-World Policies.” At the AIRoA exhibition booth, we plan to showcase live demonstrations of models currently under development using Franka + YUBI and ELEY.
Through paper presentations, the workshop, and live robot demonstrations, AIRoA will showcase a broad range of its research and development efforts — from large-scale real-world robot data collection to robot foundation models, improving the generalization and robustness of VLA models, and deployment on real robots.
Research Summaries
See Less, Specify More: Visual Evidence Budgets for Generalizable VLAs
When faced with new tasks, robots often have to determine both what to do and where to look based only on brief instructions and visually complex camera observations. S2 addresses this challenge by providing more specific guidance at each stage of a task while directing the model toward the visual information that is relevant for execution. In real-robot evaluations, S2 enabled more reliable performance on unfamiliar tasks and in visually cluttered environments.
YUBI: Yielding Universal Bidigital Interface for Bimanual Dexterous Manipulation at Scale
Conventional handheld data-collection devices such as UMI (Universal Manipulation Interface) typically use relatively bulky pistol-grip designs, which can make fine-grained bimanual manipulation difficult. YUBI is a lightweight, finger-worn gripper that opens and closes in response to the movements of a person’s finger while using VR-based tracking to record its motion.
Using YUBI, the team built a large-scale dataset comprising 8,434 hours of data, 1.20 million episodes, and 119 tasks. The study further demonstrates that a single model trained on this dataset can be deployed across multiple robot platforms.
CounterAlign: Counterfactual Supervision for Vision-Language-Action Models
Robots are typically trained from successful demonstrations, meaning that they see many examples of what they should do but have little opportunity to encounter examples of what they should not do. CounterAlign creates such instruction-inconsistent examples by pairing existing demonstrations with alternative instructions or actions, enabling policy optimization through offline inverse reinforcement learning without collecting additional robot data.
The method improves the robot’s ability to follow instructions even when object positions or tasks change, with effectiveness demonstrated in both simulation and real-robot experiments.
Nov 12 : AIRoA-Organized Workshop
The UMI Arena: From Handheld Devices to Real-World Policies
AIRoA will organize the CoRL 2026 workshop, “The UMI Arena: From Handheld Devices to Real-World Policies.”
Robot-free data collection using handheld interfaces such as UMI has rapidly expanded, enabling large-scale collection of real-world manipulation data. However, comparing the policies learned from these datasets — or the devices used to collect them — remains difficult because research groups often use different robots, tasks, environments, and evaluation protocols.
The UMI Arena aims to create a shared evaluation ground for this growing ecosystem. Policies are evaluated on common, real-robot hardware, performing the same tasks under the same conditions, while handheld data-collection devices can be compared through hands-on demonstrations.
The workshop consists of three tracks:
- Policy Competition — teams train policies using the released dataset, which are then evaluated by the organizers on shared real robots and common tasks.
- Device Exhibition — teams showcase UMI-style data-collection devices, allowing attendees to try and compare different approaches firsthand.
- Contributed Papers — researchers present work addressing open challenges in areas such as data collection, cross-embodiment transfer, VLA models, large-scale datasets, benchmarking, and real-world robot learning.
By bringing together data-collection devices, datasets, policy learning, and real-robot evaluation within a shared framework, the workshop aims to make the progress in real-world manipulation learning more measurable, comparable, and reproducible.
For details and registration for the competition, device exhibition, and paper track, please visit the dedicated workshop website:
Nov 9-10 : Demo in the Official CoRL Demo Area
YUBI: Yielding Universal Bidigital Interface from Portable Data Collection to Cross-Embodiment VLA Deployment
AIRoA’s demo proposal has been accepted to CoRL 2026, and we will conduct an interactive demonstration in the official CoRL Demo Area on November 9 and 10.
In this demo, visitors will collect robot training data themselves using YUBI, a handheld data-collection device. The collected demonstrations are converted into a LeRobot-format dataset and made available for viewing on Hugging Face. The data are then used for training, after which the resulting policy is run on a Franka Duo on site.
The demo allows visitors to experience the full pipeline firsthand — from collecting data themselves to training and execution on a real robot.
AIRoA Booth
AIRoA will also exhibit at CoRL 2026. At our booth, we plan to demonstrate models currently under development using Franka + YUBI and ELEY, alongside introductions to AIRoA’s latest research and development activities, including large-scale real-world robot data collection and curation and the development of robotic foundation models.
Through live robot demonstrations, visitors will be able to see how AIRoA brings together data collection, model development, and hardware in real-world robotic systems.
The booth will also provide an opportunity to connect with researchers, engineers, and others from around the world who are interested in research collaboration and partnership with AIRoA. We are looking forward to welcoming you at the AIRoA booth!
Where to Find AIRoA at CoRL 2026
Exhibition Booth
Booth 210, Griffin Hall, Level 2
- Nov 9: 12:00–18:30
- Nov 10: 10:00–18:30
- Nov 11: 8:30–13:00
AIRoA will showcase its latest research and development activities, including live demonstrations using Franka + YUBI and ELEY.
Demo Booth
Lone Star 7, Lone Star Ballroom, Level 3
- Nov 9: 12:00–18:30
- Nov 10: 10:00–18:30
An interactive demonstration based on AIRoA’s accepted YUBI demo proposal will be presented in the official CoRL Demo Area.
Workshop – The UMI Arena
Talks: Rooms 303–304, Level 3
Posters & Hardware Demos: Lone Star Ballroom, Level 3
- Nov 12: 14:00–18:00
AIRoA will organize a workshop bringing together research on handheld data-collection devices, datasets, policy learning, and real-robot evaluation.
*All programs will take place at JW Marriott Austin, the main venue for CoRL 2026, located at 110 E 2nd St, Austin, TX.
Featured Research Papers
See Less, Specify More: Visual Evidence Budgets for Generalizable VLAs
Authors
Yueh-Hua Wu · Tatsuya Matsushima · Kei Ota
Abstract
Generalization remains a central bottleneck for vision-language-action (VLA) models: under distractors, appearance shifts, and semantically similar tasks, the policy must often infer local execution details from coarse instructions while also deciding which parts of the image matter for control. We present S2 (See Less, Specify More), a framework for improving VLA generalization by training the executor under a cleaner interface.
Specify More preserves the original instruction as a stable high-level goal while relabeling each trajectory into refined trajectory- and subtask-level language that disambiguates the current execution mode. Unlike native attention, See Less imposes an explicit visual evidence budget, training the executor to act from task-sufficient evidence rather than unconstrained visual context, without any region or mask annotation.
This interface lets the executor follow detailed guidance without relying on distracting visual patches or resolving avoidable ambiguity on its own, and it remains compatible with off-the-shelf VLM planners through in-context learning. Across our main evaluation settings, S2 improves overall generalization metrics by changing the executor’s learning problem: coarse instructions induce avoidable supervision aliasing, goal-preserving local guidance outperforms instruction replacement in our main ablations, and explicit evidence budgeting reduces dependence on broad visual context beyond efficiency considerations.
Across eight real-robot tasks on TX-G2 (an AgiBot G2-compatible variant) and HSR, S2 raises mean subtask success from 54.2% to 79.0% over pi0.5. Together, these results suggest that VLA generalization improves when the executor is trained to act from informative local guidance and task-sufficient visual evidence, rather than recovering both from weak supervision.
Project Page: https://s2.airoa.io
YUBI: Yielding Universal Bidigital Interface for Bimanual Dexterous Manipulation at Scale
Authors
Takehiko Ohkawa · Jumpei Arima · Yuki Noguchi · Masatoshi Tateno · Makoto Sugiura · Takuya Okubo · Kengo Ikeuchi · Yuma Shin · Hiroki Nishizawa · Naoaki Kanazawa · Yuki Wakayama · Daiki Fukunaga · Koshi Makihara · Tomohiro Motoda · Floris Erich · Yukiyasu Domae · Tatsuya Matsushima · Yoshihiro Okumatsu · Kei Ota
Abstract
We introduce Yielding Universal Bidigital Interface (YUBI), a finger-aligned gripper designed to enable intuitive, ergonomic, and scalable data collection for bimanual dexterous manipulation. While handheld data collection systems such as Universal Manipulation Interface (UMI) enable affordable data collection, their bulky pistol-grip designs can pose ergonomic and usability challenges for fine-grained, dexterous manipulation tasks.
To address this, YUBI presents a distinct design principle: yielding, finger-driven actuation that directly maps human finger movements to gripper jaw motion. Using the YUBI devices, we set up a data collection system with integrated VR-based 6 DoF tracking of the gripper, ensuring high-fidelity trajectory data acquisition.
We curate a UMI-based dataset of unprecedented scale: 8,434 hours across 1.20M episodes and 119 tasks. Experiments show that YUBI offers advantages over the UMI gripper in versatility for complex bimanual tasks, dexterity, and operational efficiency.
A single policy trained on the YUBI dataset transfers across multiple bimanual robots (UR, Franka, and ELEY) simply by mounting the gripper on each platform, confirming that the collected data are directly executable as policy supervision. We release the gripper hardware, data-collection software, and dataset as one integrated stack, offering the open community a reproducible path to large-scale data acquisition for advancing robotic foundation models.
Project Page: https://yubi.airoa.io
CounterAlign: Counterfactual Supervision for Vision-Language-Action Models
Authors: Haru Kondoh · Kei Ota · Asako Kanezaki · Yueh-Hua Wu
Abstract
Vision-Language-Action (VLA) models are typically trained with behavior cloning (BC) on expert demonstrations. However, BC provides only positive supervision for expert actions, without explicit negative supervision indicating which actions are instruction-inconsistent or otherwise inappropriate. Reinforcement learning (RL) can provide such corrective signals, but often relies on externally specified rewards or curated non-expert data, both of which are costly to obtain in robotics.
We show that offline RL for VLA models need not rely on curated non-expert trajectories: successful expert demonstrations alone can be transformed into dense corrective supervision through instruction relabeling. Specifically, by pairing expert actions with mismatched alternative instructions, we synthesize counterfactual instruction-observation-action tuples from the dataset and combine them with adversarial discriminator training to learn an instruction-grounded reward model for offline RL, without collecting additional rollouts or annotations.
On the robustness-focused LIBERO-PRO benchmark, our method improves robustness to object position and task perturbations over a strong state-of-the-art baseline. It also outperforms competitive baselines in real-robot experiments on the TX-G2 (compatible with AGIBot G2).
More broadly, our results suggest that, for data-constrained VLA learning, extracting denser supervision from each demonstration can complement collecting additional data.
Project Page: https://counteralign.airoa.io
About CoRL 2026
The Conference on Robot Learning (CoRL) is an annual international conference focused on the intersection of robotics and machine learning. Since 2017, it has brought together researchers, engineers, and industry experts to share the latest research and explore new possibilities for intelligent robots.
Now in its 10th iteration, CoRL continues to be an important event for advancing robot learning and shaping the future of robotics. CoRL 2026 will be held from November 9–12, 2026, in Austin, Texas, USA, with the main conference taking place November 9–11 and workshops on November 12.