AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data

Yueh-Hua Wu, Kei Ota, Hanlong Li, Hayata Yuasa, Jason Wang, Kengo Ikeuchi, Kensuke Chiyoma, Makoto Sugiura, Masafumi Takahashi, Masatoshi Tateno, Naoaki Kanazawa, Nicola Scianca, Petr Khrapchenkov, Satoshi Otsubo, Takehiko Ohkawa, Takuya Okubo, Tatsuya Matsushima, Yuki Noguchi, Yuma Shin, Zixuan Zhong

November 2026

Colour sorting on the bimanual Franka while the bins are moved mid-run.

TL;DR Pretrain a π0.5-based policy on 16,151 hours of handheld-gripper demonstrations with no robot data at all, then post-train it per task on a few hours of handheld data. On ten long bimanual tasks at a site outside its data, A-0 averages 78% success against 27% for the best of four post-trained VLA baselines.

What they do

AIRoA-0 (A-0) learns bimanual manipulation from 16,151 hours of demonstrations recorded with YUBI, an open-source handheld gripper. It uses no teleoperation, RL, simulation or on-robot corrections. The policy sees only two wrist cameras and is tested at a site where none of its data was recorded. The dataset, pre-trained weights and post-training recipe are released.

How it works

Anchored actions. A π0.5 backbone predicts gripper poses relative to the pose at observation time, so handheld and robot trajectories share coordinates and one set of weights drives different arms.
Pre-training. Demonstrations carry task and subtask instructions (10.65M subtask segments). A visual budget gate learns, without annotation, which wrist-image tokens to keep, and an auxiliary loss predicts the current subtask.
Post-training. A stage head tracks progress through long tasks (EMA-smoothed at run time) and a memory attends over the last ten proprioceptive states. One recipe serves every task, in under four hours on eight H200s.
Non-uniform streaming decoding. Two buffered action chunks at noise levels (0.1, 1) keep motion continuous between chunks; a query takes about 60 ms on an H200.

Results

Ten tasks on a bimanual Franka (15 trials each, partial credit): A-0 averages 78% against 27% (π0.5), 23% (XR-1), 16% (GR00T N1.7) and 14% (MolmoAct2), each baseline with a per-task hyperparameter search. A-0 is best on every task, e.g. 73% on the robot vacuum, where no baseline exceeds 14%.
When a person pushes the bins aside mid-placement, the arm moves to the correct bin in one continuous motion.
Post-training loss follows a power law in pre-training hours (R² = 0.98); each doubling lowers it by about 2.4%.

Limitations

The paper has no limitations section. Each task still needs handheld post-training data: 3-6 h for six tasks, up to 25.8 h for earbuds, the weakest task at 30%. Only robots carrying the YUBI gripper and wrist cameras can run the policy, and runs on TX-G2, OpenArm and UR are shown without success rates.

Videos

Parts kitting, the orange pyramid and swapping a robot vacuum's mop pads.

Letter blocks, the cup pyramid and towel folding.

Figures

From Figure 1 (A-0 in overview): towel folding on the bimanual Franka. Handheld demonstrations, seen through their wr... — From Figure 1 (A-0 in overview): towel folding on the bimanual Franka. Handheld demonstrations, seen through their wrist cameras, are the only data collected; the robot mounts the same gripper and cameras.

From Figure 8: where the visual budget gate puts its weight on one of 25 subtasks. The tile shows the left and right ... — From Figure 8: where the visual budget gate puts its weight on one of 25 subtasks. The tile shows the left and right wrist views the model receives. The contours trace the smoothed keep probability at three levels, and the view fades where the gate gives it less weight.

From Figure 3: one frame of a two-minute autonomous run of the robot-vacuum task, in which a new mop pad is pressed o... — From Figure 3: one frame of a two-minute autonomous run of the robot-vacuum task, in which a new mop pad is pressed onto its mount. In the full run the robot removes both used pads, attaches two new ones and pulls out the brush from a single instruction.

From Figure 5: A-0 sorting toy bolts and nuts by colour on the TX-G2 humanoid, with the same weights used on the Franka.

Abstract

AIRoA-0 (A-0) is a bimanual manipulation system learned from 16,151 hours of handheld demonstrations, with no robot data collected at any stage. The demonstrations are recorded with YUBI, our open-source handheld gripper, which lets people record manipulation data anywhere with no robot present. From a single instruction, A-0 sorts more than a dozen small parts into their numbered cells and changes the mop pads and brush of a robot vacuum; among other tasks, it unzips a pencil case and folds a towel. It performs these tasks using only a camera on each wrist, with no head or scene camera, at a site where none of its training data was recorded. Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations. We collect no teleoperation data and use no reinforcement learning or on-robot corrections, and the robot appears only at test time. The system learns which regions of each wrist view matter for the task, predicts the stage the task has reached, and decodes actions as a continuous stream. A-0 is embodiment-agnostic: it runs on robots that mount the YUBI gripper, such as a bimanual Franka, TX-G2, OpenArm and UR. We are releasing the dataset, to our knowledge the largest open dataset of handheld-gripper demonstrations with task and subtask instructions, together with the pre-trained weights and the post-training recipe. Videos of A-0 on every task are on the project page, https://a-0.airoa.io. Anyone can build the gripper, record a task of their own and post-train the model on it in under four hours on a single node of eight H200 GPUs.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong