Inside Fe₀

In embodied AI, it is a widely held belief that cross-embodiment data—a diverse mixture encompassing both heterogeneous robot data and egocentric human videos—enhances robot models. Yet, this consensus remains an opaque black box: we know it works, but we rarely understand exactly what this broad spectrum of data is teaching the robot.

To demystify this, we introduce Fe₀, an embodied foundation model. Rather than scaling data to chase benchmarks, we designed a rigorous controlled experiment. We deliberately constrained our in-domain teleoperation data to the minimum: just 3 hours of trajectories featuring only a single, atomic pick-and-place primitive, where 95% of the data involves moving a single object to a single receptacle from a highly restricted vocabulary. By establishing this extreme baseline, we guarantee that any zero-shot generalization Fe₀ exhibits beyond in-domain grasping is unequivocally powered by cross-embodiment learning.

But how do we measure this transfer scientifically? Instead of ad-hoc testing, we constructed a systematic L1–L5 capability framework—spanning from basic visual robustness and semantic grounding, up to complex relational reasoning, compositional planning, and novel action execution. We evaluate the model through both rigorous open-loop trajectory scoring and extensive closed-loop real-world trials on the IRON-R01-1.11 robot to unpack the true anatomy of cross-embodiment data.

We reveal a clear capability boundary: while heterogeneous data effectively teaches the model visual semantics and basic task logic, it falters at mid-level relational and sequential reasoning, and hits a hard ceiling when transferring precise physical execution. By systematically evaluating exactly where cross-embodiment data succeeds and where it fails, we illuminate what this heterogeneous experience actually teaches an embodied foundation model, and what it takes to push beyond current limits.

The cross-embodiment pool is deliberately diverse, spanning human activity and heterogeneous robots with dexterous hands and grippers to cover broad scenes, tasks, and action primitives. The target-embodiment teleoperation anchor, however, is intentionally minimal. It is restricted to the IRON-R01-1.11 humanoid robot and its action vocabulary is dominated by a single primitive: pick-and-place.

The scale disparity between the two is massive. The cross-embodiment pool is over three hundred times larger in terms of episodes, nearly a thousand times larger in frame count, and over fifteen hundred times larger in total duration compared to the target anchor. This extreme asymmetry makes the transfer analysis interpretable. When a trial moves beyond the anchor’s restricted object set, receptacle set, task logic, or action primitive, it directly tests whether knowledge from the broader pool can be transferred on the humanoid body.

Zooming into the target-embodiment anchor makes its constraints concrete, and reveals just how narrow the in-domain objects and containers really are.

Extreme simplicity in scale and action primitives. The dataset contains merely ~3 hours of demonstrations with an average duration of just 6.1 seconds per episode. Crucially, the physical action space is strictly confined to a single, highly uniform primitive: basic pick-and-place. Every demonstration strictly follows a rigid approach → grasp → lift → transport → release kinematic sequence.

Minimal task complexity and semantic diversity. Furthermore, the task complexity and semantic diversity are heavily restricted. An overwhelming 95% of the trajectories represent the most atomic manipulation scenario: transferring a single object to a single receptacle. The remaining 5% of the data involves placing two objects onto different tiers of a single shelf, and is strictly limited to only three object categories (apple, banana, and bread). Across the entire dataset, the visual and semantic scope is tightly bounded to 20 common household items and 6 fixed target locations (such as boxes, plates, or handing over to a person) as shown below, uniformly driven by highly templated language instructions.

Objects

breadbananabowlapplewhiteboard erasermarkerplastic bottlespoonmouseblockcokeclothcuptissuesscrewdrivertennis ballpaper cupplush toyfruitegg

Containers

boxshelfplatebasketpersonbowl

To make cross-embodiment transfer measurable, Fe₀ is evaluated through a Multi-Axis Decomposition Framework rather than a single aggregate metric. The five axes are not arbitrary capability buckets—each one is designed to evaluate a specific type of reusable knowledge that the heterogeneous data pool can plausibly supply, mapping cleanly from perception to action: visual robustness and semantic grounding at the input side, relational reasoning and compositional planning in the middle, and novel action prediction at the output side.

For every layer we collected a matched set of teleoperated trajectories on the IRON-R01-1.11 to serve as ground truth. The model is then evaluated open-loop: along each recorded trajectory we query the model frame by frame and score it by the MSE between its predicted actions and the corresponding teleoperated actions. Unlike a closed-loop trial on the real robot, this protocol scores every frame from the ground-truth state rather than from where the model would have ended up—the success of preceding steps is never enforced. As a result, scores on long-horizon, multi-step layers (most visibly L4 Compositional Planning) may diverge from real-robot outcomes, where errors normally compound across steps.

Evaluating transfer across the L1–L5 framework reveals a clear capability boundary in cross-embodiment learning. While scaling heterogeneous data effectively solves visual semantics and basic task logic, it falters at mid-level relational reasoning and hits a hard ceiling when transferring precise physical execution. This capability drop-off exposes deep structural bottlenecks: the lossy nature of multimodal representations, the challenge of overcoming behavioral biases inherent in teleoperation data, and the fundamental limits of learning physical interaction from raw trajectories. True physical generalization requires new representational and learning paradigms, not just more data.

Yizhuo Li^*, Xiaoyu Zhang^*, Yuying Ge^*†, Teng Wang, Boyu Chen, Yi Chen, Bo Liu, Feng Qiu, Yuguo Gan, Hui Zhou, Yixiao Ge

• ^*Co-first authors.
• ^†Project lead.

@article{li2026fe0,
  author = {Yizhuo Li and Xiaoyu Zhang and Yuying Ge and Teng Wang and Boyu Chen and Yi Chen and Bo Liu and Feng Qiu and Yuguo Gan and Hui Zhou and Yixiao Ge},
  title = {Inside Fe0: What Cross-Embodiment Data Teaches an Embodied Foundation Model},
  journal = {XPENG Robotics Blog},
  year = {2026},
  note = {https://xpeng-robotics.github.io/fe0/},
}