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

TL;DR

AIRoA-0 (A-0) fine-tunes a π0.5-shape VLA on 16,151 hours of bimanual handheld YUBI-gripper demonstrations (10.65M subtask segments) with no robot data whatsoever — no teleoperation, no RL, no on-robot corrections. The same gripper mounts as the robot end-effector, making images and actions embodiment-invariant at collection time (UMI-lineage). A-0 reports 78% average success on ten bimanual Franka tasks at a site where none of the training data was recorded, vs ≤27% for four other VLA models post-trained on the same demonstrations, and the same checkpoint runs on bimanual Franka, TX-G2 semi-humanoid, OpenArm, and UR via gripper-swap. Post-training on a new task takes <4h on 8×H200. Dataset, weights, and post-training code promised at release.

Key claims

78% average success on ten held-site bimanual tasks (15 trials each, partial credit for partial completion); every other model tested maxes at 14–27% [§05 Results].
Trained on 16,151 hours of bimanual YUBI handheld-gripper demonstrations (10.65M subtask segments, 5.5 s avg each, 1.74B time steps at 30 Hz, two wrist cameras only — no head or scene camera) [§01 Data].
Action loss after post-training follows a power law in hours of pre-training data; post-training on the same demonstrations does not compensate for less pre-training data beyond the first ~1000 updates [§04 Scaling].
Buffer of 2 action chunks: new chunks are denoised in a single step from noise level 1.0 to 0.1, then refined to clean prediction on the next query — amortizes denoising across queries so only two passes of the action network per query are needed [tweet note, §02 System].
Last 10 gripper states as proprioceptive history are compressed into 4 tokens before being fed to the policy [tweet note].
The system’s four parts — data processing, pre-training (visual-evidence budget + two-level language), post-training (adds task-stage prediction + recent-motion memory), and inference harness (EMA-smoothed stage + non-uniform streaming decoding) — are co-designed so each part’s outputs are the next’s inputs [§02 System].
Pre-training teaches the model, without visual annotation, which regions of each wrist image matter for the task — the “where to look” mechanism Martin flagged as unverified [§02 System].
Demonstrators record with YUBI; the same gripper mounts on different arms (Franka, TX-G2, OpenArm, UR) as end-effector, so image views and action commands match at collection and deployment [§01 Data, §03 Demonstrations].
<4h post-training on a new task on one node of 8×H200 GPUs [§00 header].
Dataset, pre-trained weights, and post-training recipe pledged for release at the time of this writeup [§06 Release].

Method

A-0 is a π0.5-shape VLA fine-tuned exclusively on handheld bimanual UMI-lineage data. The YUBI gripper serves as both the demonstrator’s handheld tool and the robot’s end-effector, so the two wrist cameras see what the robot will see and the recorded gripper pose is what the robot will command. The pre-training phase imposes a visual-evidence budget and consumes two-level language annotations (task + subtask) per segment; the model learns an implicit “where to look” selector without explicit visual supervision. Post-training adds two components the handheld corpus lacks: a task-stage predictor supervised by the subtask boundaries, and a 4-token compression of the last 10 gripper states as recent-motion memory.

The inference harness runs an asynchronous rolling buffer of 2 action chunks. When a new query arrives, the next chunk is initialized from Gaussian noise at level 1.0 and denoised in a single network pass to noise level 0.1; the previously partially-denoised chunk is then refined to its clean prediction. This amortizes the ten-step denoising process across queries so each query only requires two forward passes of the action network, streaming actions out non-uniformly while an EMA smooths the predicted task stage.

Results

78% average success across 10 bimanual Franka tasks, 15 trials per task, partial credit; A-0 wins every task individually [§05 Results].
Four comparison VLAs post-trained on the same per-task handheld demonstrations top out at 14–27% average; each entry is each model’s best over action horizons {4, 6, 8, 10, 12, 14, 16}, multiple action frequencies, and {60k, 80k, 100k} updates [§05 Results].
Evaluation at a site none of the training data was recorded at — tests the generalization story of handheld collection rather than scene memorization [§05 Results].
Pre-training data scaling (6.25% → 100%): held-out action loss decreases power-law in hours; the gap between shares is preserved after post-training on the same 16 task sets [§04 Scaling].
Same weights deployed on bimanual Franka (3× real-time in demos), TX-G2 semi-humanoid (5×), OpenArm (3×), and UR (2×) via gripper swap, driven by a single language instruction given once at episode start [§03 Demonstrations].

Why it’s interesting

A-0 is the cleanest filed datapoint yet for the “make-the-collection-tool-embodiment-invariant” branch of Human-to-Robot Retargeting pushed to the no-robot-data at all extreme. It sits directly between Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1 (100K-h UMI pre-training but still 7,200 h teleop post-training) and HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone HiFi-UMI (3 mm fidelity, 2 K-h, matches teleop baseline on three VLA backbones) by dropping the teleop leg entirely at 16 K-h — an operating point neither occupied. It is also an existence proof against Pixel-UMI: Learning Environment-Aware Whole-Body Manipulation from Sensorized Human Demonstrations Pixel-UMI’s claim that human-UMI pre-training needs DAgger-style on-robot refinement to clear 87% whole-body task progress: A-0 reports 78% without any on-robot data, suggesting the DAgger leg is task-regime-dependent rather than universal. The two-pass async denoising buffer is a concrete entry on the chunk-latency axis alongside πR²: Reactive Real-time Flow Policies πR² (latency-adaptive staircase + slow/fast conditioning split) and Real-Time Robot Execution with Masked Action Chunking REMAC (LoRA + prefix-mask training) — distinct because the two passes amortize full ten-step denoising across queries rather than restructuring either the model or the training recipe.