E0.5 takes a 3D foundation model, built to reconstruct the scene in front of it, and teaches it to reconstruct the future: the depth, point maps and cameras of the scene to come. Then it acts on that geometry.
Most world-action models (WAMs) inherit their world knowledge from video generators or vision-language models, so geometry is either implicit or an optional side input. E0.5 takes the opposite route: it is built directly on a 3D foundation model and trains that backbone end-to-end to predict the future 3D state (depth, point maps and camera parameters at the end of the action chunk) jointly with the actions that lead there.
Without any embodied-data pretraining, E0.5 reaches 89.90% success on clean RoboTwin 2.0 scenes and 50.09% on randomized scenes with a 2.32B-parameter policy. Predicting future geometry adds 9.43 points of average success over no reconstruction and 7.93 points over reconstructing the current state. After pretraining on more than 160k AgiBotWorld-Beta trajectories, mean real-world success rises by about 15 points on AgileX Piper and about 7 points on the Zerith H1 humanoid.
89.90%RoboTwin 2.0 clean success, 50 tasks
+9.4 ptsaverage success from predicting future geometry
2.32Bpolicy parameters, no embodied pretraining
+15 / +7pts real-world mean SR from pretraining · Piper / H1
01 / OverviewNative 3D backbone, explicit geometric futures
Think in geometry. Reconstruct the future.
01 / GEOMETRY
See the scene as geometry
The head and wrist camera images and the instruction go into a 3D foundation model: a backbone built to recover depth, points and camera poses, not pixels or words.
02 / FUTURE
Reconstruct the future in 3D
The same backbone reconstructs the scene that is about to happen, as depth, point maps and camera poses. This is a training signal only; nothing extra runs at inference.
03 / ACT
Act on that geometry
An action expert reads those geometry features layer by layer and generates the next chunk of robot motion.
Most world-action models
Start from a video model or a VLM.
Image + instructionVideo model or VLMFuture pixels, or no future at all
Geometry stays implicit, or is bolted on as an optional extra input.
E0.5
Start from a 3D foundation model and reconstruct the future in 3D.
Now · head camerafuturePredicted depthPredicted point cloud
Real E0.5 output on the RoboTwin task stack_blocks_two: holding the green block, it predicts that the arm will soon be over the red block. The actions come from the same features.
02 / Real worldAgileX Piper · Zerith H1
On real robots. One model, two embodiments.
E0.5 is post-trained on a dual-arm AgileX Piper and on the Zerith H1 dual-arm humanoid, with a head camera and two wrist cameras. Each video shows the third-person view alongside the three policy camera streams.
Real-world success: E0.5 vs. π0.5
π0.5 · no pretrainE0.5 · no pretrainπ0.5 · pretrainedE0.5 · pretrained
10 trials per policy, task, setting and platform. "No pretrain" compares E0.5 against Gemma-initialized π0.5; "pretrained" compares E0.5 pretrained on AgiBotWorld-Beta against the official π0.5-base checkpoint. OOD uses larger trash sets with unseen objects, unfamiliar chargers and cables, and unseen mouse poses. Completion score: charger body 0.25 + full cable 0.75; trash scored k/N.
AgileX Piper · E0.5 mean SR (ID + OOD)70.0% → 85.0%with embodied pretraining
Zerith H1 · E0.5 mean SR (ID + OOD)81.7% → 88.3%with embodied pretraining
03 / MethodNative 3D world-action model
One 3D backbone that reconstructs the future. Geometry teaches it; actions use it.
During training, the backbone has to predict the scene's geometry at the end of each action chunk, a short way into the future. Its features therefore encode how the scene is about to change, and the action expert reads those features at every layer. At deployment only the action path runs.
InputThree camera viewsheadleft wristright wrist
InputLanguage instruction
“Stack the green block on the red block.”
Backbone3D foundation model
Starts from a model pretrained for 3D reconstruction and is trained end to end, not frozen.
reconstructs
every layer
Training onlyFuture geometryDepthPoint mapCamera
A supervision signal. Removed at deployment, so it costs nothing at inference.
OutputAction expert
Flow matching, conditioned on the backbone's geometry features.
Next action chunk →
SimulationRoboTwin 2.0: 50 demos per task, no embodied pretraining
PretrainingAgiBotWorld-Beta: 160k trajectories, about 2,650 h
Real robots200 demos per task on AgileX Piper and Zerith H1
Pick any of the 50 RoboTwin tasks and compare the depth and point clouds E0.5 predicts for its head and wrist cameras with what actually happened.
adjust_bottle1 / 50
Each video is the first training trajectory of its task, not picked by quality; predictions are shown without alignment. Depth uses scene-normalized units. Point colors come from the real RGB frame for display only; E0.5 does not generate images.
Why it matters
Predicting the future is what helps.
Under the same training budget, supervising the backbone with future geometry raises mean RoboTwin success from 46.53% to 55.96%. Reconstructing only the current scene reaches 48.03%.
What the backbone is trained to predict
05 / SimulationRoboTwin 2.0 · clean → randomized
Watch it act in simulation. Strong results without embodied pretraining.
Trained on 50 clean demonstrations per task, then evaluated in clean scenes and in domain-randomized scenes it never saw during training.
Clips are selected successful episodes shown from the head camera; they illustrate behavior, while the numbers above are success rates over all trials. Twelve tasks have no randomized clip and turn_switch has none yet.
Data efficiency
No robot-data pretraining needed.
With no embodied pretraining at all, E0.5 averages 70.0% across clean and randomized scenes. That is higher than π0.5, 4D-WAM and GigaBrain-0.7, which pretrain on 10,000 to 37,000 hours of robot data.
Success vs. embodied pretraining data
x = hours of embodied pretraining (0 = none). y = mean of clean and randomized success. Hover a point for details.
Full comparison with 15 published methods
Success rate on RoboTwin 2.0
Success rate (%), macro-averaged over 50 tasks. Params = policy parameters in billions. Pretrain = embodied pretraining data (– = not reported; ✗ = none). Baseline numbers as reported in the E0.5 technical report.
06 / CitationBibTeX
@article{lai2026e05,
title = {E0.5: From Reconstructive to Future-Predictive 3D Foundations
for Robotic World-Action Modeling},
author = {Lai, Jianhao and Zhan, Zhihao and Liao, Zhiquan and Li, Qichang and
Chen, Shubin and Liu, Zhiyuan and Lei, Hongsen and Lin, Jianman and
Xu, Zhihua and Lin, Xiaoxin and Gong, Ke and Gu, Baiyue and
Chen, Shiming and He, Junlin and Yang, Longqi and Xiao, Junrong and
Luo, Haolin and Su, Xiaojuan and Li, Jianfeng and Luo, Yu and
Sun, Xu and Zhang, Likui and Zhang, Jusheng and Lyu, Qinhan and
Chen, Tianshui and Lin, Liang and Wang, Guangrun},
journal = {arXiv preprint arXiv:2610.36011},
year = {2026}
}
E0.5 / FUTURE GEOMETRY PREDICTION
Anticipate the future. Act in the world.
From future physical structure to embodied action. E0.5 is part of X-Era.AI's work on world-action models.
