WANDA: One Demo Generates Infinite Training Data for Mobile Robots
From a single human demo, WANDA creates thousands of varied training trajectories.
Training robots for open-world mobile manipulation typically requires vast amounts of human teleoperation data—a costly and labor-intensive bottleneck. A new paper from researchers Lingxiao Guo, Huanyu Li, and Guanya Shi introduces WANDA, a synthetic data engine that maximizes the utility of a single human demonstration. WANDA first reconstructs the scene as Gaussian splats and captures the robot-object interaction trajectory from a single RGBD observation. Then it rearranges contact-rich segments into new spatial configurations using whole-body motion planning, effectively multiplying the original demo into countless unique training trajectories.
To enhance robustness in long-horizon tasks, WANDA applies Corrective State Expansion, which increases diversity in robot and object states throughout the manipulation. For cross-environment generalization, trajectories are synthesized on diverse 3D worlds generated from everyday photos. The final training observations are photorealistic composites of robot/object meshes with Gaussian splatted backgrounds. Experiments show that policies trained with WANDA generalize well across environments and even to different robot morphologies (zero-shot on another mobile manipulator). This dramatically reduces the data bottleneck for deploying generalist robots in the real world.
- WANDA reconstructs scene and robot-object interactions from one RGBD demo using Gaussian splats.
- It generates thousands of new trajectories via whole-body motion planning and Corrective State Expansion.
- Policies trained with WANDA achieve cross-environment and cross-embodiment generalization in real-world tasks.
Why It Matters
One human demo can now train a robot for complex mobile tasks—cutting data costs by orders of magnitude.