The X-Era AI Lab team published a paper on 26 October 2026 describing E0.5, a world-action model whose primary backbone is a 3D geometry model rather than a video or vision-language one. On the dual-arm AgileX Piper and the Zerith H1 humanoid it reached mean success rates of 85.0% and 88.3%, against 71.7% and 61.7% for π0.5. The code is on GitHub under an MIT licence, with weights and data on Hugging Face.
Key takeaways
- AgileX Piper: 85.0% mean success against 71.7% for π0.5, a gap of 13.3 percentage points
- Zerith H1: 88.3% against 61.7% for π0.5, a gap of nearly 27 percentage points
- RoboTwin 2.0 without any embodied pretraining: 89.90% on clean scenes and 50.09% on randomized ones
- The policy core holds 2.32 billion parameters against roughly 3.3 billion in π0.5
- Pretraining ran on 160,010 trajectories from AgiBotWorld-Beta, about 2,645 hours of recordings
Where this starts
A robot has to pack a charger into a box — together with its entire cable. The body is rigid and always looks much the same, but the cable is soft and settles differently every time. In the X-Era tests this task scored 0.25 points for enclosing the body and 0.75 for the whole cord.
To manage it, the model has to predict where the cable will be a second from now — not how it will look, but where it will physically sit in space.
And here is the problem: a typical model learns to predict the next video frame, where a change in the cable's shape is tangled up with changes in lighting, shadow and reflection.
E0.5 is built to address exactly that, by predicting future geometry instead of a future picture.
What exactly did the X-Era AI team build?
E0.5 sits on VGGT-Omega, an off-the-shelf 3D reconstruction model recovering depth and scene layout from a handful of images. In most competing systems such a model is at best an extra source of features, with a video or vision-language model as the primary backbone. Here it is the other way round — the geometric backbone does the computation, and its weights are trained.
The second change concerns the training target. Alongside the movement sequence, the model predicts the scene's future 3D state: depth maps?depth maps: images in which every pixel stores its distance from the camera rather than a colour, point maps and camera parameters for the frame matching the final step of the planned action. Action and predicted geometry describe the same stretch of time.
The movement itself comes from a separate module based on the DiT architecture from WAN2.2, coupled to the backbone in a mixture-of-transformers?mixture-of-transformers: an arrangement in which several specialised transformers share one data stream, each handling a different part of the task arrangement. Language instructions enter through a frozen T5 encoder. The model plans 32 steps ahead using 10 denoising?denoising: the gradual removal of noise from a random sample, the process by which a diffusion model arrives at a finished output steps.
Per the paper, switching from reconstructing the current scene to predicting the future one raised mean success by 7.93 percentage points, and by 9.43 points against a variant with no geometric reconstruction.
The predicted geometry is purely a training signal — at run time the model neither needs nor receives it.
What do these names mean?
The paper leans on several terms the numbers depend on.
- World-action model (WAM) — a model learning two things at once: which movement to make and how the surroundings will look afterwards. The equivalent of a player planning a move together with the board layout that follows.
- VGGT-Omega — an off-the-shelf 3D reconstruction model recovering scene geometry from images. A starting point, like a pretrained backbone in image classification.
- Point map — a record assigning every pixel a position in space. Comparable to a photo in which each point knows its distance from the camera.
- Action chunk — a batch of movements planned at once, here 32 steps of 14 joint and gripper values. Instead of deciding frame by frame, the model issues the whole sequence.
- π0.5 (pi-zero-5) — a vision-language-action model from Physical Intelligence, a widely used reference point in robotic manipulation.
Together these add up to one idea: E0.5 turns a model that reconstructs what is visible into one predicting what is about to happen in space, tied directly to movement.
Why not just add geometry to a video model?
Today's standard puts a video or vision-language model in the primary backbone role, with 3D information entering as extra features or an auxiliary training target. That works and raises scores, but geometry stays an add-on to a representation created to describe appearance.
The difference shows up outside the training distribution?outside the training distribution: in situations the model never saw during training — unfamiliar objects, scenes or conditions. A model trained to predict pixels models lighting, texture and background alongside object shape. With an unfamiliar cable or a different table layout, that knowledge stops helping. Per the paper, in the evaluated out-of-distribution setting E0.5 reached roughly thirty times the accuracy of the compared video-prediction-based models.
There is a second, practical problem: who trains the backbone. The related VGA approach also uses a native geometric backbone, but reconstructs the current scene and adapts with LoRA, because full-parameter training degraded its results. E0.5 unfreezes the whole backbone and trains it jointly with the action module — without that, there is no move from reconstruction to prediction.
What do the real-robot tests show?
The team tested three tasks on two platforms: placing a mouse on a pad, packing a charger with its cable, and dropping trash into a bin. Each came in an easy variant matching the training data and a hard one with unfamiliar objects, at 120 trials per policy.
The largest difference is on the humanoid. On the Zerith H1, pretrained E0.5 averaged 88.3% success against 61.7% for π0.5-base, with a clean 100% on the easy variant of all three tasks. On the dual-arm Piper the lead is smaller but clear: 85.0% against 71.7%.
Cable packing proved hardest. With unfamiliar chargers, E0.5 reached 80% on the Piper and 50% on the H1, while π0.5-base stopped at 40% and 30%.
In simulation the picture is mixed. On RoboTwin 2.0, E0.5 without embodied pretraining leads π0.5 on average (70.00 against 58.33) but trails PatchWAM and MWA-Phoenix, near 79 points each. Its strength is the ratio of result to model size and training cost, not the raw score.
| Platform | E0.5 | π0.5 |
|---|---|---|
| AgileX Piper (physical robot) | 85.0% | 71.7% |
| Zerith H1 (physical robot) | 88.3% | 61.7% |
| RoboTwin 2.0 (simulation, no embodied pretraining) | 70.00 | 58.33 |
Is it working yet?
Yes, as open research code — not as a product. The X-EraAI/e05 repository was created on 27 October 2026 under an MIT licence. Hugging Face hosts the RoboTwin policy, an embodied-pretrained VGGT backbone for real-robot adaptation, and a processed dataset of roughly 378 GiB.
The limits are stated plainly. The MIT licence covers repository code only and does not change third-party terms — VGGT-Omega is noncommercial, research-only. The reference environment is Linux with eight GPUs, while the original RoboTwin training ran on 32. The repository ships no policies for the Piper and H1, only the backbone for fine-tuning them.
Why does it matter?
The result challenges the default assumption that the best foundation for robot control is a model trained on video or image-text pairs. If spatial structure transfers better than appearance, the cost calculation shifts: a smaller model achieves more on less data. Open weights let that claim be tested outside the team that made it.
What comes next?
- The real-robot numbers rest on 120 trials per policy across three tabletop tasks — confirming the lead needs independent repeats on a wider task set
- The repository notes the full 50-task RoboTwin evaluation was not rerun during code validation, so the paper's figures await outside reproduction
- VGGT-Omega's noncommercial terms remain a barrier to production deployment, despite the MIT licence on the code
Sources
- arXiv — E0.5: From Reconstructive to Future-Predictive 3D Foundations for Robotic World-Action Modeling
- GitHub — X-EraAI/e05
- X-Era.AI — E0.5 · Future-Predictive 3D World-Action Model
- Hugging Face — X-EraAI/e05-robotwin
