WANDA — Worlds in One Demo

01 — Real-world demonstrations

Five tasks, one demo each

Five long-horizon mobile-manipulation tasks on Agibot G1, each learned from a single human demonstration. Browse autonomous real-world rollouts across spatial configurations and scenes.

02 — Interactive viewer

Explore the 4D worlds interactively

Open each task scene as an interactive replay with the Gaussian-splat background, moving objects, and robot trajectory. Source and generated demonstration videos are shown below the task scenes.

Source scenes

Background Gaussian splats are reconstructed with MAtCha from the single demonstration.

Generated scenes

Background Gaussian splats are generated by Marble from one input photograph.

Objects

Object geometry is reconstructed using a refined BundleSDF pipeline.

03 — Cross-embodiment

One demo, a different robot

WANDA reuses the same reconstructed world across robot embodiments. From one Agibot G1 Drop Trash demonstration, it generates Linearbot trajectories and deploys the policy zero-shot on a physical Linearbot with different kinematics and camera poses.

Agibot G1 versus Linearbot: 7-dof vs 6-dof arms, 2-dof vs 1-dof waist, non-holonomic vs holonomic base, different camera placements

Not a lookalike robot

Arms7-dof vs 6-dof
Waist2-dof vs 1-dof
Basenon-holonomic vs holonomic
Cameras3 each — very different poses

WANDA does not copy joint motion. It re-plans whole-body trajectories for the new robot in the same world, then renders observations from that robot's camera viewpoints.

Zero-shot · Linearbot Spatial configuration 1.

Zero-shot · Linearbot Spatial configuration 2.

04 — The idea

One demo contains a world

A demonstration is more than pixel-action pairs to imitate. As the robot moves, its cameras observe the scene: one demo contains enough signal to rebuild the world around the task.

Instead of treating the demo only as action labels, WANDA uses it to recover the scene, object motion, and contact-rich interaction. That recovered world becomes the substrate for generating more training data.

Robot data has been priced in human hours. WANDA prices it in compute.

Teleoperation scales with human effort. WANDA keeps the human input fixed at one demonstration, then uses reconstruction, generation, planning, and rendering to scale the dataset.

05 — Method

One demonstration, diverse data

WANDA turns a single demonstration into diverse mobile-manipulation data through reconstruction, diversification, and corrective expansion.

Reconstruct

Recover the scene representation, object geometry, and 6D object trajectories from the source demonstration.

Diversify

Instantiate the recovered interaction across sampled spatial configurations and photo-generated scenes.

Expand

Apply Corrective State Expansion to sample off-demonstration states and train recovery behavior.

06 — Results

One demo → 54.8% progress

π_0.5 fine-tuned only on WANDA-generated data, evaluated on physical robots and long-horizon simulation benchmarks. Every WANDA result starts from one demonstration.

Real world · progress score over 10 trials · Agibot G1 · five long-horizon tasks

Corrective State Expansion improves average real-world progress from 15.7% to 54.8% by training the policy on recovery states it encounters during execution.

Data efficiency · BEHAVIOR Challenge Q-score · 3-minute task

One demo expanded by WANDA outperforms π_0.5 trained on 50 teleoperated demos. On Bigym, one demo is close to training on the full source set.

Cost · wall-clock hours per 1 h of data · shorter is better

Factorized rendering generates one data-hour in about 3 GPU-hours, using less GPU time than MoMaGen and avoiding repeated teleoperation for each data-hour.

07 — Abstract

Abstract

Learning open-world mobile manipulation needs vast data for spatial generalization, long-horizon robustness, and scene generalization. We maximize the value of a single demonstration with a synthetic data engine.

Teleoperation and UMI require substantial human effort at scale. We introduce WANDA: learning open-World mobile mANipulation from one demonstration via a synthetic DAta engine. WANDA reconstructs background Gaussian splats and robot–object interaction trajectories from source RGBD observations as a world substrate, rearranges contact-rich interaction segments into extensive spatial configurations, and chains them with whole-body motion planning. Corrective State Expansion increases state diversity for long-horizon robustness, while trajectories are synthesized on diverse generated 3D worlds from everyday photos for cross-environment generalization. Photo-realistic observations are produced by compositing rendered robot and object meshes with Gaussian splatting backgrounds. Across simulation and real-world tasks, policies trained with WANDA achieve long-horizon robustness, broad spatial generalization, and cross-environment generalization from a single real demonstration — and naturally support cross-embodiment data generation, validated by zero-shot deployment on a distinct mobile manipulator.

08 — Limitations

Limitations

WANDA handles rigid and articulated objects — reconstruction tracks 6D part poses and replays them — so cloth and fluids are replayed as demonstrated rather than physically modeled. Because WANDA is a multi-stage engine, reconstruction or planning errors can propagate downstream.

Contact-rich interaction is still provided by the single source demonstration. Future systems could replace full robot teleoperation with lighter sources such as UMI-style grippers or human video.

09 — Reference

Citation

@article{guo2026wanda,
  title   = {Worlds in One Demo: A Synthetic Data Engine for
             Learning Open-World Mobile Manipulation},
  author  = {Guo, Lingxiao and Li, Huanyu and Shi, Guanya},
  year    = {2026}
}