InfiNoVA

Abstract

Vision-Language-Action (VLA) policies often rely strongly on the camera viewpoints seen during training, causing substantial performance degradation when deployed from un- seen perspectives. Collecting demonstrations from sufficiently diverse physical viewpoints is expensive and still provides only sparse coverage of the viewpoint space. We introduce InfiNoVA, a data-augmentation framework that converts syn- chronized multi-camera demonstrations into a dense distri- bution of geometrically consistent training views. InfiNoVA reconstructs each manipulation trajectory as a time-varying 3D Gaussian representation and renders novel observations from sampled camera poses while preserving the original state–action correspondence. This explicit scene representation improves frame-level fidelity and temporal consistency while reducing task-critical hallucinations observed in generative novel-view synthesis. Across four real-world manipulation tasks, policies trained with InfiNoVA achieve 5.4× higher average success under unseen randomized viewpoints than both VISTA-based augmentation and the unaugmented policy. InfiNoVA further achieves 1.7× higher success than training directly on all five physical camera views. These results show that dense, geometrically grounded viewpoint augmentation provides a practical route toward camera-robust robot policies without modifying the underlying policy architecture.

Method

InfiNoVA takes synchronized multi-camera robot demonstrations and reconstructs each manipulation trajectory as a time-varying 3D Gaussian scene representation. From this explicit geometric model, new camera viewpoints are sampled and rendered into additional training observations, while the original state–action labels from the demonstration are preserved unchanged. Because rendering is grounded in explicit 3D geometry rather than generative synthesis, the resulting views maintain high frame-level fidelity and temporal consistency, avoiding the task-relevant hallucinations that can arise from purely generative view-synthesis approaches. The resulting dense set of geometrically consistent viewpoints is then used to train policies that are more robust to unseen camera placements at deployment time.

Rendering Quality

Hunyuan

GT

Hunyuan

DepthSplat

GT

DepthSplat

GLD

GT

GLD

VISTA

GT

VISTA

InfiNoVA

GT

InfiNoVA (Ours)

InfiNoVA

VISTA

InfiNoVA (Ours) vs VISTA

Compared to generative and depth-based baselines such as Hunyuan, DepthSplat, GLD, and VISTA, novel views rendered by InfiNoVA show noticeably higher frame-level fidelity to the ground-truth observation, with sharper object boundaries and fewer structural artifacts around the robot arm and manipulated objects. Several baselines introduce visible hallucinations or geometric distortion in task-relevant regions when rendering from a shifted viewpoint, which is especially problematic since these are exactly the regions a downstream policy must attend to. Because InfiNoVA renders from an explicit, time-varying 3D Gaussian reconstruction of the scene rather than synthesizing pixels generatively, it better preserves scene geometry and temporal consistency across frames, which is consistent with the substantially higher downstream policy success rates reported for InfiNoVA-augmented training data.

Rendering time is also an important practical consideration for scaling augmentation to large demonstration datasets. GLD produces visually competitive renderings but is substantially slower to render than InfiNoVA, making it considerably less practical for generating the large volume of synthetic viewpoints required for policy training. InfiNoVA's explicit 3D Gaussian representation allows novel views to be rendered efficiently, making it a more scalable choice for producing dense viewpoint augmentation across full demonstration datasets.

Averaged across tasks, InfiNoVA is approximately 3.6× faster than VISTA and 82.8× faster than GLD, while also providing higher rendering fidelity and temporal consistency, making it the strongest quality–efficiency trade-off among the evaluated methods.

A further practical limitation of DepthSplat and GLD is that both are pose-dependent methods, requiring an explicit target camera pose as input to condition the rendering process. In contrast, InfiNoVA, along with Hunyuan and VISTA, is pose-free, directly synthesizing novel views without requiring the target viewpoint to be known or specified beforehand. This pose-free property makes InfiNoVA better suited to augmentation pipelines where diverse novel viewpoints must be generated without manually specifying camera poses for every rendered view.

Novel View Rendering

Each clip above shows the same manipulation trajectory rendered from a different camera viewpoint using InfiNoVA's 3D Gaussian scene representation. The robot, objects, and motion stay sharp and consistent across all viewpoints, even ones never seen by a physical camera during data collection. This shows that the rendered views are realistic and reliable enough to be used as additional training data for the robot's policy.

Robot Experiments

When the SmolVLA policy is trained on a single fixed camera viewpoint, it performs well on a pick-and-place task as long as it is evaluated from that same training viewpoint. However, when the camera is moved to a randomized test viewpoint not seen during training, the same policy fails horribly, highlighting how sensitive standard VLA policies are to the camera viewpoint used at test time.

Across all five randomized viewpoints—spanning Pick&Place, Stack, Sort, and Sweep—the base policy and the VISTA-augmented policy both exhibit severe degradation, frequently failing to localize or grasp the target object once the camera pose deviates from the training distribution. This indicates that both baselines overfit to the specific viewpoint statistics seen during training rather than learning a viewpoint-invariant visuomotor representation. In contrast, the policy trained on InfiNoVA-augmented data continues to localize, grasp, and manipulate the target objects successfully across every task and viewpoint, showing that its robustness to viewpoint shift generalizes across task types rather than being specific to a single manipulation skill.

These videos show the InfiNoVA policy evaluated from the same randomized camera view across all four tasks.

Results

For each of the four manipulation tasks, we evaluate every policy over 100 episodes with the test-time camera pose randomized on each episode, and report the resulting task success rate.

We evaluate InfiNoVA across four real-world manipulation tasks, measuring policy success rate under unseen, randomized camera viewpoints at test time. Policies trained with InfiNoVA achieve 5.4× higher average success under these unseen viewpoints compared to both VISTA-based augmentation and an unaugmented baseline policy.

Across all four tasks, the unaugmented baseline and the VISTA-augmented policy both collapse under randomized viewpoints, achieving only 5–12% success, well below the reference success rate obtained when the baseline is instead evaluated from its original training viewpoint (52–76%). This confirms that the failure under viewpoint shift is not due to task difficulty, but to poor generalization across camera poses. The policy trained on InfiNoVA-augmented data is far more robust to this shift, reaching 38–44% success across Pick&Place, Stack, Sweep, and Sort, consistently several times higher than both baselines under the same randomized viewpoint evaluation.

Notably, InfiNoVA also outperforms policies trained directly on all five physical camera views, achieving 1.7× higher success, showing that dense, geometrically consistent synthetic viewpoints can be more effective than simply collecting more real camera coverage.

Citation

@article{infinoVA2026,
  title={InfiNoVA},
  author={Anonymous Authors},
  year={2026}
}