We conduct a systematic simulation study to examine how robot augmentation scales in terms of transfer, generalization, and robustness. While prior work has shown that augmenting demonstrations from a source robot to a known target enables zero-shot transfer, our goal is to investigate whether robot augmentation provides broader benefits when scaled across multiple target robots

Generalization to Unseen Robots

We evaluate performance on robots that were not used for augmentation. Specifically, we compare three settings: training only on the source robot (“No Augmentation”), augmenting demonstrations to a single additional robot that is not the target robot (“1 Augmented Robot”), and augmenting to all robots besides the held-out robot (“N–1 Augmented Robots”).

All policies are evaluated on an unseen robot that is not part of the augmentation set. We see that the ``Source-Only'' policy performs poorly, as expected. ``1+1'' policies generalize moderately well to some robots, but ``N-1'' achieves substantially higher success across all four target robots. This suggests that robot augmentation promotes embodiment-agnostic visual representations and spatial reasoning, and generalization to unseen embodiments improves as we increase the diversity of robot augmentations.

TODO: Add more videos

Robustness to Visual Perturbations

We also ask the question: "Can robot augmentation improve robustness on the original robot?" We evaluate on the original source robot under visual perturbations (lighting shifts and occlusions). Similar to above, we compare 3 policies: training only on the source robot ("Source-Only"), augmenting to another robot ("Source + Target (1+1)") and augmenting to N robots ("N").

The figure shows performance under test-time perturbations to the source robot's environment: (1) lighting shifts that introduce shadows, and (2) occlusions from randomly placed black rectangles. While "Source-Only" policies degrade significantly, both "1+1" and "N" are more robust, with "N" consistently achieving the highest success. This suggests that robot augmentation enhances robustness even on the original embodiment by encouraging the policy to focus on task-relevant structure rather than incidental features such as arm texture or lighting cues.

Robustness Demonstrations

In real experiments, we evaluate whether large-scale augmentation can also benefit pre-trained foundation models. We consider two state-of-the-art generalist policies—OpenVLA and $\pi_0$—and fine-tune them using OXE-Aug.

For evaluation, we use tasks from the Bridge dataset, originally collected on a WidowX robot, and test on two embodiments:

1. A Franka arm with its default parallel-jaw gripper ("Franka"), which corresponds to one of the augmented robots in OXE-Aug, and
2. A Franka arm with a custom modified Robotiq gripper ("Robotiq++"), which features colored pads to simulate an unseen embodiment.

We evaluate 4 manipulation tasks. For each base model, we compare "Source-Only," "Target-Only," and "N."

Each policy is evaluated over 10 trials per task. Both OpenVLA-OFT and $\pi_0$'s performances are relatively low when fine-tuned only on the original Bridge data, especially on Franka, due to the visual domain shift from the black WidowX gripper to the white Franka one. Fine-tuning on OXE-Aug significantly improves cross-embodiment performance: "Target-Only" improves success across all tasks, and "N," which incorporates more diverse robot augmentations, yields the highest success overall. On average, "N" improves performance by 24% for OpenVLA-OFT and 45% for $\pi_0$. On the novel Robotiq++ embodiment, fine-tuned policies reach 75% (OpenVLA-OFT) and 82% ($\pi_0$) average success, demonstrating strong generalization.

Task Videos