Handheld Data and Human Video Point to Different Paths for Scaling Robot Learning
AIRoA-0 and Liber-0 each reduce dependence on robot-collected demonstrations, but they solve different parts of the problem: one preserves action fidelity through handheld tools, while the other uses broader human experience only after building bridges to robot control.
By Felix Park · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human engineering credentials or firsthand experience.
Key points
- AIRoA-0 reports a 78% average success rate across its ten-task evaluation using wrist-mounted cameras and a model trained from handheld demonstrations rather than robot data; its stated comparison baseline reached 27% after post-training on the same task demonstrations.
Sources: S1
- Liber-0 is described as pretraining on more than 2.2 million hours of human experience, but its approach also relies on glove motion recordings, robot demonstrations and embodiment information to connect visual observation to executable action.
Sources: S2
- The central constraint is not just data volume. Ordinary human video can leave hand motion, depth, contact and force ambiguous; handheld and glove systems seek to make the action signal more observable without requiring a robot for every recording.
The scale question is really an observability question
Two developments offer a useful correction to the common idea that robot learning will scale simply by collecting more video. AIRoA-0 is a bimanual manipulation system trained on 16,151 hours of handheld demonstrations, with no robot data collected at any stage, according to its authors. LiberAI’s Liber-0, meanwhile, is described as a World Action Model pretrained on more than 2.2 million hours of human experience, much of it large-scale activity video. Both projects seek to loosen the bottleneck created by robot demonstrations, which require hardware and operator time. But their data strategies start from different answers to a harder question: what can the learning system actually see about the action that caused a physical outcome?
AIRoA-0 favors a directly measurable action channel
AIRoA-0’s answer is to put a gripper in the demonstrator’s hand. Its YUBI handheld gripper records manipulation without a robot present, while retaining a tool form that can be mounted on several robot platforms. The authors say the system is embodiment-agnostic across robots that mount YUBI, including bimanual Franka, TX-G2, OpenArm and UR systems. Rather than depending on a head or scene camera, the deployed system uses one camera on each wrist. It learns relevant regions in those wrist views, estimates task stage, and produces actions as a continuous stream.
Sources: S1
The reported evaluation matters because it moves beyond a claim of broad pretraining. At a site where none of the training data was recorded, AIRoA-0 handled tasks including sorting small parts into numbered cells, changing a robot vacuum’s mop pads and brush, unzipping a pencil case, and folding a towel. Across ten tasks, its average success rate was 78%, versus 27% for the strongest of four other vision-language-action models that were post-trained on the same task demonstrations. The authors also report no teleoperation data, reinforcement learning, or on-robot corrections, and say the robot appears only at test time.
Sources: S1
The system’s tradeoff is visible in its design. Wrist views are close to the work and can expose object-tool relationships that a distant camera may miss, but they do not guarantee a complete scene model. Occlusion, clutter outside the camera frame, or state changes away from the gripper can still make the input incomplete. AIRoA-0’s stage prediction and attention to wrist-view regions are mechanisms intended to support a decision from those limited views; the supplied abstract does not establish how the system responds when the critical state remains unseen.
Sources: S1
Sources: S1
Liber-0 treats human video as broad context, not sufficient control data
Liber-0 takes a wider-data route, but the account of its design explicitly acknowledges the limits of watching. The Robotics & Automation News report says ordinary video may not reveal precise finger motion, grip firmness, wrist rotation, or the movement sequence behind a visible result. It identifies visual viewpoint mismatch, incomplete action supervision caused by factors such as occlusion and depth-estimation error, and the embodiment gap between human and robot bodies. These are communication limits between the recorded human action and the robot policy: an image can convey that a drawer opened without fully specifying which forces, contacts, and motions made it open.
Sources: S2
LiberAI’s proposed bridge is multimodal. Its lightweight gloves record hand poses and trajectories while a first-person camera and two wrist-mounted cameras capture the activity; robot demonstrations supply examples of actions machines can execute. The company also uses an embodiment representation that includes human and robot physical characteristics. This is distinct from AIRoA-0’s route. AIRoA-0 tries to make demonstrations action-compatible through a common handheld gripper. Liber-0 attempts to learn across differing bodies by representing those differences and combining broad video with smaller, more action-specific data sources.
The reported results support the narrower claim that glove data can complement robot data, not that human video alone solves robot manipulation. In an eight-task evaluation, a policy using 25 percent of the full Robot demonstration pool and 100 percent of the Glove pool recorded a 75.6 progress score; a policy using 100 percent of the Robot pool recorded 72.8. In another reported comparison, robot demonstrations limited to pick-and-place produced a 2.2 progress score on other manipulation, while adding glove demonstrations covering non-pick-and-place actions raised the score to 47.5. Broader glove data reached 54.6, near the 56.3 reported for the Teleop All reference.
Sources: S2
The original comparison: scale needs an action bridge
Inference: the two systems suggest that the scalable unit of robotics data is not a video clip alone, but a record that links observation to a feasible action for a particular body. AIRoA-0 emphasizes that link by collecting demonstrations through the tool it later mounts on robots. Liber-0 seeks to create the link through glove trajectories, robot data, aligned cameras, and embodiment-conditioned learning. Their shared dependency is action observability; their difference is whether compatibility is designed into collection hardware or learned from heterogeneous records.
This distinction also explains why their headline results should not be treated as a ranking. AIRoA-0 reports task success across its own ten-task evaluation and comparison models. Liber-0’s supplied results include progress scores for separate training mixtures, platform comparisons, and a 36% success rate across a separate suite of 60 zero-shot tasks. The workloads, metrics, data mixtures, robot arrangements, and operating conditions differ. Neither supplied account provides a common benchmark that would justify declaring one approach superior overall.
What to watch when the input is incomplete
The practical test is whether these systems can identify when perception is insufficient rather than continuing with a plausible but unsafe action. AIRoA-0’s wrist-only deployment makes robustness to hidden task state, camera blockage, and changes outside the immediate manipulation region especially important. Liber-0’s human-video strategy makes robustness to missing depth, occluded fingers, and unmeasured contact especially important. LiberAI itself says contact and force remain important dimensions for future development, underscoring that visual coverage and trajectories do not fully describe manipulation.
Evidence that could change this assessment would include common evaluations across robot bodies and unseen environments, with clearly matched task definitions and failure reporting. Particularly informative results would separate failures caused by perception, embodiment mismatch, contact uncertainty, and long-horizon task planning. For AIRoA-0, evidence on when wrist views fail to reveal needed state would test the limits of its compact sensing setup. For Liber-0, ablations separating ordinary video, glove trajectories, robot demonstrations, and embodiment representation would show which bridge is carrying transfer. The larger opportunity is real in both approaches: collect useful experience without putting a robot into every data-collection session. The unresolved engineering question is how much of that experience remains actionable when the robot’s view, body, or physical contact differs from the one in the record.
Why it matters
Robot learning may become cheaper to scale if physical demonstrations can be recorded without robot hardware. But broader data only becomes useful control knowledge when the system can resolve the action, body, and contact information that video leaves uncertain. AIRoA-0 and Liber-0 illustrate two different engineering paths toward that bridge, with different sensing assumptions and different risks when observations are incomplete.
Sources
- AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data — arXiv Robotics ·
- From human experience to robot action: LiberAI launches Liber-0 world-action model — Robotics & Automation News ·