Liber-0
Emergent Zero-Shot Capabilities in Multi-Step Manipulation from 2.2 Million-Hour Human Experience
[Video: Liber-0 overview]

Beyond reasoning, human intelligence relies on intuition to interact with the physical world—coordinating perception and movement, regulating force, and adapting to feedback without explicit deliberation. While AI has made remarkable progress in reasoning, developing comparable intuition for manipulation remains a fundamental challenge.

Traditionally, robot manipulation relies on demonstrations collected for specific robots and tasks, an approach that is costly and hard to scale [1]. Human experience offers a much broader range of physical interactions, much of it already captured as internet-scale video. Recent studies have demonstrated the value of learning from such data, with emerging evidence of favorable scaling trends [2], [3], [4].

However, learning from human video is less straightforward than learning from text in large language models (LLMs). Human experience does not directly translate into executable robot actions, owing to three gaps:

• Embodiment Gap. Humans and robots differ in morphology, kinematics, and ranges of motion. The same manipulation objective can therefore require different movements across embodiments.
• Action-Supervision Gap. Hand poses and trajectories inferred from video can be noisy or incomplete due to occlusions, inaccurate depth, and estimation errors, limiting their reliability as supervision for precise manipulation [5].
• Visual Gap. Human recordings and robot observations differ in camera placement, viewpoint, and the appearance of hands and bodies, creating a visual distribution shift at deployment [6].

These gaps motivate a central question: How can we turn large-scale physical experience from human activities into generalizable action capabilities?

To address these challenges, we introduce Liber-0, a World Action Model (WAM) pretrained on over 2.2 million hours of human experience. Our approach combines large-scale visual learning with action grounding and embodiment alignment, following three complementary design principles (Figure 1):

• Learn embodiment as context. A Unified Embodiment Representation (Section 2) encodes human and robot morphology and kinematics, conditioning action prediction on the executing body to support transfer across embodiments.
• Ground human video with measured action. LiberSense Glove system (Section 3.1) provides precise hand motions paired with human-perspective video, serving as action anchors that connect large-scale visual experience with executable actions through joint video-action learning.
• Bring robot observation and motion closer to human demonstrations. Our robot platform, LEO (Section 3.2), combines dexterous hands, articulated wrists and waist, and aligned camera configurations to reduce visual and kinematic differences between human experience and robot execution. Such a design further facilitates cross-embodiment transfer.

Our experiments demonstrate the effectiveness of this approach at scale. Glove recordings exhibit particularly strong scaling in fine-grained manipulation, outperforming 200 times more Web Human Data on both PnP and Non-PnP benchmarks (Section 4.1). Combined with limited robot supervision, Glove data enables performance exceeding Robot-only training with 75% fewer robot demonstrations (Section 5.1), while also transferring manipulation skills absent from the robot-demonstration set (Section 5.2). Moreover, increasing human morphology diversity improves transfer even at a fixed data volume (Section 5.3). Together, these findings support our approach to action grounding and embodiment alignment, with the full staged pretraining recipe achieving the strongest overall benchmark performance (Section 4.3).

Surprisingly, after pretraining, Liber-0 demonstrates task-level zero-shot manipulation (Section 4.2), executing tasks without robot demonstrations collected specifically for them. On 60 random tasks proposed by GPT-6, evaluated over 600 trials, Liber-0 achieves an overall success rate of 36.0%, including pick-and-place (PnP), short non-pick-and-place (Non-PnP), and challenging multi-step Non-PnP tasks (see below). These results suggest that action-grounded human experience offers a scalable foundation for generalizable robot manipulation.

Selected zero-shot demonstrations  [More evaluation results \nearrow]
PnP: Cup placement; Pouch placement; Card placement.
Non-PnP: Object reorientation; Paper folding; Tissue extraction.
Challenging Multi-Step: Packing and storage; Seasoning and covering; Closing and wiping.
[Nine embedded videos]

Beyond these zero-shot tasks, Liber-0's strong pretrained manipulation capabilities enable rapid adaptation to application-specific workflows, such as Boba tea preparation as shown in Video 1.

We begin by describing how Liber-0 addresses the challenges in learning from human experience: the lack of reliable action supervision and the differences in embodiment between humans and robots. As illustrated in Figure 2, our approach combines action-grounded pretraining with embodiment-conditioned learning, complemented by semantic co-training.

Action-Grounded Pretraining. We adopt world-action modeling [7] to jointly predict future visual frames and actions. LiberSense Glove recordings provide action anchors by pairing human-perspective video with precisely measured hand motions, bridging the gap between large-scale human video and executable actions. Joint prediction allows visual learning to scale beyond action-annotated data while grounding action learning in measured motion. We implement this framework using a Mixture-of-Transformers architecture with dedicated language, vision, and action experts [8].

Embodiment as Context. The same manipulation objective may require different movements across bodies due to differences in morphology and kinematics. We introduce a Unified Embodiment Representation that encodes human and robot geometry and kinematics in a shared space. Conditioning action prediction on this representation allows Liber-0 to learn how movements vary across embodiments, supporting transfer from diverse human demonstrations to closely aligned robots.

Semantic Co-Training. Beyond physical dynamics, manipulation requires understanding task semantics, spatial relationships, and execution progress [9]. We therefore co-train Liber-0 on visual question answering (VQA), image segmentation, image editing, and video understanding. These objectives complement world-action learning with semantic and temporal understanding while preserving broad visual capabilities. For multi-step tasks, Liber-0 can optionally integrate with a reasoning agent, such as GPT-6 Astra, to plan intermediate goals and guide execution.

Apart from the learning framework, effective human-to-robot transfer also requires alignment at the hardware level. We develop two complementary systems: LiberSense Glove provides reliable action supervision when recording human demonstrations, while LEO Robot reduces the visual and kinematic differences between human experience and robot execution.

LiberSense Glove is a lightweight wearable system for collecting human manipulation data with directly measured finger motion. Unlike rigid exoskeleton-based designs [10], it is only 0.3 mm thick and weighs 131 g. This design minimizes interference with finger movements and hand–object interaction, preserving the natural manipulation patterns we want the model to learn. Its skin-toned exterior also maintains visual similarity to bare hands commonly seen in egocentric video.

The system synchronizes measured hand poses and trajectories with video from a first-person camera and two wrist-mounted cameras. By preserving the visual perspective and embodiment of human video while providing precise action supervision, these recordings serve as action anchors for large-scale human-data pretraining.

Importantly, collection requires no robot on site, enabling low-cost, scalable demonstrations across diverse human morphologies, objects, tasks, and environments.

LEO is a robot platform we designed to minimize the visual and kinematic gap between human demonstrations and robot execution. It combines dexterous hands for fine-grained manipulation, a slim forearm structure with a human-like wrist, and an articulated waist that can reach various angles and heights.

LEO features three degrees of rotational freedom at the wrist, enabling a human-like range of wrist motion without moving the forearm. This enables precise rotational manipulation and hand placement in confined spaces. Its slim, human-like forearms further reduce collision risks, while an articulated waist extends the reachable workspace and enables the robot to reproduce bending and reaching motions commonly found in human demonstrations. Together, these designs make a broader range of human movements directly executable on LEO.

Visual alignment is also built into data collection and deployment. The first-person and wrist-camera viewpoints used in glove collection closely match LEO's chest- and wrist-camera configuration. As shown in Video 2, human and robot observations capture hand–object interactions from similar perspectives, further reducing the distribution shift between pretraining and inference.

With the learning framework and hardware alignment in place, we turn to the central question of this work: can scaling human experience translate into generalizable robot manipulation skills? Our experiments reveal two key findings. First, action-grounded glove data exhibits particularly strong scaling, especially for fine-grained manipulation. Second, Liber-0 generalizes beyond predefined benchmarks to tasks without task-specific robot demonstrations, including challenging multi-step manipulation. We present these results before examining the training strategy and evaluation protocols.

We first investigate how different data sources contribute to robot manipulation as data scales. We separately pretrain models on five sources: Glove, Cross-Embodiment Robot Data, OpenEgo, VendorEgo, and Web Human Data (Web in Figure 5), using 1, 2, 5, and 10 kh of data per source. For the Web, we additionally evaluate 100 kh and 2,000 kh. All models are post-trained on the same robot dataset and evaluated on 60 PnP and 16 Non-PnP tasks, using average progress scores as metrics.

Figure 5 summarizes the experimental results, from which we draw five key findings.

All data sources show favorable scaling trends. Every evaluated source improves over the baseline without source-specific pretraining, which scores 39.3 on PnP and 34.1 on Non-PnP. Despite local fluctuations, performance generally improves as data increases, demonstrating that diverse sources of physical experience can contribute to robot manipulation.

PnP and Non-PnP exhibit distinct scaling patterns. For sources other than Glove, additional data yields substantially smaller gains on Non-PnP than on PnP. Glove, by contrast, maintains strong scaling across both groups: increasing from 1 kh to 10 kh improves PnP by 14.7 points and Non-PnP by 13.1 points. This suggests that broad experience transfers more readily to object selection and relocation, while fine-grained manipulation depends more strongly on reliable, transferable motion supervision.

Glove delivers the strongest performance, particularly on Non-PnP. On PnP, Glove and Cross-Embodiment Robot Data exhibit comparable scaling trends and generally outperform the other sources. On Non-PnP, Glove achieves both higher performance and stronger scaling gains, highlighting the effectiveness of action-grounded human data for precise and sustained manipulation.

Web-scale volume alone does not close the transfer gap. Scaling Web Human Data to 2,000 kh improves its scores to 64.6 on PnP and 51.2 on Non-PnP, both below the results obtained with only 10 kh of Glove recordings. Despite a 200-fold difference in duration, Glove transfers more effectively to robot manipulation. This contrast underscores the importance of combining broad visual experience with reliable action supervision and close human–robot alignment.

The full Liber-0 recipe achieves the highest overall performance. Combining large-scale human video with action-grounded data through staged pretraining, Liber-0 reaches 83.6 on PnP and 73.8 on Non-PnP after robot post-training, substantially exceeding every single-source configuration. This result supports our strategy of combining complementary data sources rather than relying on a single source at increasing scale.

Together, these results reveal complementary roles for visual coverage and motion supervision. Large-scale human video provides broad physical experience, while Glove recordings offer particularly effective supervision for transferring that experience into executable actions. These findings motivate our staged pretraining strategy, detailed in Section 4.3.

Beyond improvements on predefined benchmarks, the ultimate test of scalable pretraining is whether learned capabilities generalize to tasks without task-specific robot demonstrations.

To evaluate this, we test Liber-0 on 60 tasks proposed by GPT-6, independently of our internal scaling benchmarks. The suite covers three levels of difficulty: 20 PnP tasks, 20 short Non-PnP manipulation tasks, and 20 challenging multi-step Non-PnP tasks. Each task is evaluated over 10 trials, yielding 600 trials in total. The evaluated checkpoint has not been trained on robot demonstrations collected for these target tasks.

Liber-0 achieves success rates of 63.5% on PnP, 34.0% on short Non-PnP, and 10.5% on challenging multi-step Non-PnP tasks, with an overall success rate of 36.0%. Notably, the model successfully executes some multi-step manipulation sequences without task-specific robot training, extending zero-shot execution beyond simple grasping and placement. Representative zero-shot trials are shown in Section 1.

We also evaluate GPT-6 Astra as a standalone robot policy on the 60 GPT-6-generated tabletop tasks, with three trials per task (180 trials in total). Astra-only achieves success rates of 8.3%, 5.0%, and 1.7% on PnP, short Non-PnP, and challenging multi-step Non-PnP tasks, respectively. Astra's average reasoning time per decision is 87.3 s, while the mean end-to-end completion time for successful tasks, including both reasoning and robot execution, is 17.8 min.

In our evaluation, Astra-only struggles with contact-rich manipulation requiring precise alignment and controlled hand–object contact. Its successful behaviors are largely confined to coarse grasping and placement that tolerate imprecise motion. In the accompanying video, for example, it repeatedly fails at the initial drawer-opening step. Long control cycles further limit timely feedback correction, consistent with the latency issues discussed in [11].

We further explore object-level generalization through an interactive online simulation application. After post-training Liber-0 on 100 hours of simulated teleoperation data, users can upload custom objects and observe manipulation in simulation without additional object-specific training. This provides an open-ended complement to our fixed real-robot benchmarks.

These results demonstrate that large-scale human experience, combined with action grounding and embodiment alignment, can support robot manipulation beyond the specific tasks represented in robot demonstrations.

Guided by the complementary strengths of different data sources, we pretrain Liber-0 on over 2.2 million hours of human experience using a three-stage strategy (Figure 6). Training progressively shifts from broad visual prediction toward action-grounded manipulation.

Stage 1: General Visual Prediction. We begin with visual prediction pretraining across diverse human activities, objects, and environments. At this stage, the model learns to predict future observations without an action-training objective. This establishes broad priors about object interactions and physical dynamics, providing a foundation for subsequent action learning.

Stage 2: Joint World-Action Pretraining. We then switch to joint prediction of future frames and actions. The training mixture combines the visual coverage of human video with motion supervision from glove and robot recordings, allowing the model to connect how an interaction unfolds with the actions that produce it. This stage brings visual prediction and action learning into a shared framework across diverse embodiments.

Stage 3: Manipulation-Focused Pretraining. In the final stage, we increase the sampling weight of Glove and teleoperation recordings. This shift places greater emphasis on precise motion and sustained interaction, building on the broad priors acquired in earlier stages to strengthen manipulation performance.

As shown in Figure 7, performance improves across successive training stages in both direct evaluation (without additional robot post-training) and evaluation after robot post-training. By Stage 3, the pretrained checkpoint reaches progress scores of 81.4 on PnP and 72.3 on Non-PnP, with robot post-training further improving these scores to 83.6 and 73.8, respectively.

Importantly, the additional benefit of robot post-training decreases as pretraining advances. At Stage 3, post-training adds only 2.2 points on PnP and 1.5 points on Non-PnP, suggesting that much of the manipulation capability present in post-training has already been acquired through pretraining. We have reason to expect that, in deployment, a properly pretrained model may significantly reduce the need for post-training.

Finally, we elaborate on the composition of the training data, as well as the evaluation benchmark used in the preceding experiments.

Training Data. Our pretraining corpus combines complementary sources of visual experience and motion supervision:

• Glove Data: Human manipulation recordings collected with LiberSense Glove, pairing measured hand poses and trajectories with synchronized first-person and wrist-camera video. The dataset covers diverse PnP and Non-PnP behaviors, human morphologies, objects, and environments.
• Cross-Embodiment Robot Data: Robot manipulation demonstrations spanning different embodiments, tasks, and environments, providing explicit robot action supervision.
• OpenEgo: Human egocentric video aggregated from public datasets, covering diverse activities and hand–object interactions.
• VendorEgo: Externally collected egocentric human manipulation recordings, offering more consistent demonstrations and motion patterns closer to robot execution.
• Web Human Data: Curated online human activity videos, providing broad coverage of objects, scenes, viewpoints, and physical interactions.

To strengthen semantic, spatial, and temporal grounding capabilities, we additionally co-train on general visual understanding data, including:

• VQA: Image question-answering and captioning examples covering object recognition, scene description, and visual relationships. Task-conditioned subtask prediction examples pair overall instructions with current and historical observations to identify the active execution step, connecting task semantics with manipulation progress.
• Image Segmentation: Pixel-level segmentation examples providing spatial supervision over object regions and boundaries, grounding semantic labels in precise visual regions.
• Image Editing: Instruction-conditioned image editing examples linking language instructions to corresponding visual transformations, providing supervision for how intended changes should appear in the scene.
• Video Understanding: Video QA, dense captioning, and temporal grounding examples covering action sequences, event order, object-state changes, and execution progress. Video QA captures temporal and interaction-level understanding, dense captions describe successive interactions and their outcomes, and temporal grounding aligns language-described actions with corresponding video intervals [12]. Together, these tasks provide supervision for tracking interaction progress over time.

Benchmarks. Our internal scaling benchmarks evaluate two complementary categories of manipulation: pick-and-place (PnP), which probes object selection and placement across varied objects and scenes, and Non-PnP, which evaluates manipulation behaviors including precise hand–object interaction, object-state control, and multi-step action coordination. These tasks require the policy to adapt its movements to object geometry and mechanical constraints and achieve the intended final object configuration.

• Pick-and-Place (PnP). The PnP benchmark includes two 30-task sets: one using the existing object inventory and another using separately procured object instances spanning both familiar and additional categories. Each set contains 40 objects across five tabletop scenes, with eight objects present in each scene. Tasks require selecting the instructed object and placing it in, on, or next to a specified target. Each task is evaluated over three trials, testing object identification, grasping, and spatially specified placement under variations in object appearance and category.
• Non-PnP Manipulation. The Non-PnP benchmark comprises 16 tasks across six action families: folding, opening, closing, pressing, pulling, and pushing. Examples include folding a towel into a storage box, closing an eyeglass case, operating a drawer, and pressing a button on a scale. Some tasks combine these interactions with grasping and placement, requiring completion of the full manipulation sequence. Each task is evaluated over six trials, giving 96 trials per model.

Evaluation Metrics and Protocol. We report progress scores on a 0–100 scale to capture partial execution alongside complete task performance. PnP scoring tracks milestones including approach, alignment, grasping, transport, release, and stable placement. Non-PnP scoring assigns task-specific weights to completed action milestones. Initial configurations are restored using reference images or videos, and model comparisons use matched instructions and scene configurations.

Notably, these progress-based benchmarks are distinct from the GPT-6-proposed zero-shot task suite in Section 4.2, which reports task success rates rather than partial progress.

The aforementioned results demonstrate that large-scale human experience can translate into robot manipulation capabilities. We now investigate what makes this transfer effective, examining how glove recordings reduce the need for robot demonstrations, enable new manipulation skills, and benefit from diversity and alignment across embodiments.

We first investigate how effectively glove recordings can substitute for robot demonstrations. We consider two settings: combining increasing amounts of glove data with limited robot demonstrations (Glove + Limited Robot Data), and training on glove recordings alone (Glove-only). Robot-only training serves as the reference (Figure 8).

All configurations use the same number of training steps and are evaluated on eight tasks using LEO equipped with WUJI2 hands, with 10 trials per task (80 trials per configuration). The Robot and Glove data pools contain equal numbers of demonstrations.

Learning with limited robot data. In the Glove + Limited Robot Data configuration, robot demonstrations are fixed at 25% of the Robot pool. Increasing Glove data from 25% to 50% and 100% raises progress from 61.2 to 68.4 and 75.6, respectively. The final configuration exceeds the 72.8 achieved by Robot-only training with the full Robot pool, while using 75% fewer robot demonstrations.

Transfer from glove recordings alone. Under Glove-only training, using 50% of the Glove pool achieves a progress score of 49.4, approaching the 50.0 obtained with 25% of the Robot pool. Increasing Glove data to 100% further raises progress to 60.3, exceeding the 50.0 achieved by Robot-only training with 25% of the Robot pool.

These results demonstrate that glove recordings can substantially reduce reliance on robot demonstrations, while supporting meaningful cross-embodiment transfer even without robot data in policy training. Together, they establish glove collection as a scalable alternative to increasing teleoperation data alone.

Beyond reducing robot data requirements, a more important question is whether glove recordings can teach robots skills absent from their robot demonstrations.

We investigate this by varying the PnP and Non-PnP coverage of the robot and glove data (Figure 9). Teleop PnP serves as the baseline, while Teleop All provides a reference trained on both PnP and Non-PnP robot demonstrations. We compare these with two mixed configurations: Teleop PnP + Glove Non-PnP, and Teleop PnP + Glove All. Each configuration is evaluated on four PnP and six Non-PnP tasks, with 10 trials per task.

Two findings emerge from Figure 9:

First, aligned glove recordings transfer manipulation behaviors beyond robot demonstrations. When robot demonstrations cover only PnP tasks, adding Glove Non-PnP raises Non-PnP progress from 2.2 to 47.5. Using Glove All further increases it to 54.6, approaching the 56.3 achieved by Teleop All. These results suggest that high-quality motion supervision, combined with visual and embodiment alignment, makes fine-grained human manipulation experience transferable to robot execution, enabling behaviors absent from the robot-demonstration set.

Second, shared behaviors anchor transfer across embodiments. Expanding Glove Non-PnP to Glove All raises PnP progress from 54.3 to 69.5 and Non-PnP progress from 47.5 to 54.6. Shared PnP behaviors provide corresponding examples of manipulation under human and robot embodiments, helping the model relate human movements to robot actions and transfer additional skills.

These findings highlight a key advantage of action-grounded human pretraining: glove data can expand the robot's skill repertoire, rather than merely improving performance on skills already demonstrated by robots. Shared behaviors further facilitate this transfer by providing common ground across embodiments.

Human morphology varies in hand geometry, body proportions, and movement patterns, even when performing the same task (Figure 10). We investigate whether greater human embodiment diversity improves the model's cross-embodiment ability, independently of the total amount of training data.

We vary the number of demonstrators while keeping the total number of glove episodes constant. Glove + Limited Robot Data combines glove recordings with a small, fixed set of robot demonstrations, while Glove-only uses glove recordings alone in policy training. Robot-only uses the corresponding robot demonstrations without glove recordings and serves as the reference. Both glove-based configurations use 1, 3, 10, 30, or 100 demonstrators. All demonstrator groups use the same collection tasks and environments. We evaluate all configurations on two PnP and four Non-PnP tasks.

As shown in Figure 11, performance improves consistently with demonstrator diversity. Increasing coverage from 1 to 100 demonstrators raises progress from 60.6 to 69.5 in the Glove + Limited Robot Data setting. The gain is more pronounced in Glove-only training, where progress increases from 12.7 to 40.1, narrowing the gap to the Robot-only reference of 51.8.

Importantly, these improvements occur without increasing the total number of glove episodes, suggesting that diversity in human embodiments provides valuable supervision beyond data volume alone. By learning how similar skills are executed across different human bodies, the model may better generalize to the target robot's morphology. This raises the possibility that sufficiently diverse human embodiments could make a closely aligned robot an interpolation target within the learned embodiment space, further reducing the need for robot-specific supervision.

The hardware designs in Section 3 aim to make the motion patterns of humans directly accessible to robot execution. To study their contribution to skill transfer, we compare LEO with a platform equipped with Marvin M6 arms from Tianji, referred to as Tianji below. We focus on tasks requiring substantial wrist rotation, precise hand placement, and coordinated waist–arm movement.

We train separate policies using the same glove recordings and equal amounts of platform-specific teleoperation data. For each task, the same operator collects demonstrations on both platforms, keeping object initial positions, action sequences, and intended movement strategies as consistent as possible. This reduces variation in operator style and task setup when comparing the two embodiments.

As illustrated in Video 5, the three tasks expose distinct movement requirements:

• Knob turning: large wrist rotations with a stable grasp. The fingers must remain engaged with the knob as the hand rotates. LEO's wrist articulation accommodates this motion with limited compensatory arm movement, allowing the policy to follow the turning pattern present in human demonstrations.
• Slipper arrangement: bending to reach objects near the floor. The robot must lower its upper body, extend its arms, and orient its fingers within a low workspace. LEO coordinates waist and arm movement to follow the human sequence of bending, reaching, grasping, and repositioning.
• Drawer opening: precise hand placement in confined spaces. The handle sits close to the tabletop, leaving little room for the hand and wrist. Humans typically approach with the palm facing forward and the wrist bent upward, then pull along the drawer's direction of travel. Reproducing this motion requires precise finger placement and dexterous wrist positioning to avoid collisions during grasping and pulling.

LEO outperforms Tianji on all three tasks, achieving an average progress score 28.3 points higher. These results illustrate the value of kinematic compatibility: when a robot can reproduce the movement patterns present in human demonstrations, the learned manipulation strategies can transfer more effectively into executable robot actions.

We present Liber-0, a World Action Model pretrained on over 2.2 million hours of human experience. By combining large-scale human video with action grounding from LiberSense Glove, embodiment-conditioned learning, and human-aligned robot hardware, our approach enables human experience to be translated into executable robot skills at scale.

Our experiments demonstrate favorable scaling across diverse data sources, with glove recordings showing particularly strong gains when scaled up. Glove data also substantially reduces the need for robot demonstrations, enables the transfer of skills absent from robot training data, and benefits from greater human morphology diversity and closer robot–human kinematic alignment. Liber-0 further demonstrates zero-shot generalization to unseen objects and tasks, including challenging multi-step manipulation. Together, these findings support action grounding and embodiment alignment as complementary principles for scaling robot intelligence from human experience.

Toward richer physical supervision. While LiberSense Glove captures precise hand motions, physical interaction also depends on contact and force, which cannot always be inferred from motion alone. Incorporating tactile and force sensing could provide richer supervision for contact-intensive manipulation, including insertion, fastening, and deformable-object handling. The challenge is to capture these signals without compromising the lightweight design that enables scalable human data collection.

Toward human experience with gloves at greater scale. Unlike robot teleoperation, glove collection does not require a robot at each site, making it possible to scale human manipulation data across increasingly diverse people, tasks, and everyday environments. Building on our existing collection framework, we envision expanding both the scale and diversity of human experience, approaching the breadth of interactions encountered in daily life. Combined with richer physical supervision, this offers a promising path toward more generalizable robot manipulation.

[1] Khazatsky et al. DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset. RSS, 2024.
[2] Dyna Robotics. Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models. Technical report, 2026.
[3] Light Origins. Light-O1: Scaling Whole-Body Intelligence with Human Action Pretraining. Technical report, 2026.
[4] Zheng et al. EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data. arXiv:2602.16710, 2026.
[5] Wang et al. DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation. RSS, 2024.
[6] Kareer et al. EgoMimic: Scaling Imitation Learning via Egocentric Video. arXiv:2410.24221, 2024.
[7] Ye et al. World Action Models are Zero-shot Policies. arXiv:2602.15922, 2026.
[8] Liang et al. Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models. TMLR, 2025.
[9] Physical Intelligence et al. \pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization. arXiv:2504.16054, 2025.
[10] Xu et al. DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation. CoRL, 2025.
[11] Yang, Khetan, and Zhang. Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy. NVIDIA technical report, 2026.
[12] Bai et al. Qwen3-VL Technical Report. arXiv:2511.21631, 2025.