Beyond Viewpoint Generalization:
What Multi-View Demonstrations Offer and How to Synthesize Them for Robot Manipulation?

Abstract

Understanding Multi-View Gains in Robot Learning

While multi-view demonstrations are known to improve robustness to camera shifts, we investigated whether they fundamentally enhance manipulation capabilities. By isolating viewpoint diversity and evaluating all policies from a single canonical camera, we ensured performance gains were not simply due to test-time view availability.

• Finding 1: Improves manipulation performance. Multi-view supervision consistently increases success rates, especially in visually diverse environments where reliance on view-specific cues is high.
• Finding 2: Transfers to VLA models. The benefits extend to large, pretrained Vision-Language-Action models, improving success rates across both seen and unseen tasks.
• Finding 3: Moderate diversity works best. Performance gains are non-monotonic; an intermediate number of added views (4-8) yields peak results, avoiding the noise of excessive appearance variation.
• Finding 4: Moderate magnitude offsets are optimal. Viewpoint shifts between ±10° and ±40° enrich geometric cues effectively, whereas larger shifts (±50°+) introduce hindering appearance changes.
• Finding 5: Scales with additional demonstrations. Increasing the number of demonstrations per added view steadily improves policy performance without early saturation.
• Finding 6: Breaks single-view saturation. Unlike single-view training, which plateaus as data increases, multi-view training continues to climb, overcoming traditional scaling ceilings.

Internal Mechanisms: Multi-view supervision reshapes visual representations to focus on manipulation-relevant regions (like the end-effector and objects) rather than the background. It also improves action head robustness and stabilizes optimization dynamics.

➜ Motivated by these insights, we propose RoboNVS to synthesize high-quality multi-view demonstrations from monocular inputs.

RoboNVS: Synthesizing Multi-View Demonstrations

To overcome the scarcity of multi-view data, we propose RoboNVS, a geometry-aware framework that synthesizes novel-view videos from a single monocular demonstration. Why Geometry Matters? Direct video generation often fails in robotics due to geometric inconsistency (e.g., distorted objects or hallucinated structures). Therefore, we explicitly incorporate 3D geometry as a strong prior.

Overview: RoboNVS reconstructs 3D geometry from monocular video, renders it under new viewpoints, and uses diffusion-based inpainting to generate physically consistent novel-view videos.

Together, these designs enable RoboNVS to generate geometrically consistent, temporally stable, and physically plausible novel-view demonstrations.

Extensive Visual Comparison

Experimental Results

Figure 1. Real-world Setup. Our evaluation environment and the three manipulation tasks: Click Bell, Pick Fruit, and Pick Lego. Each task requires precise geometric understanding of the target objects.

Figure 2. Qualitative Comparison. Visual results in real-world scenarios. RoboNVS demonstrates superior performance in both geometry reconstruction and mask completion quality compared to existing baselines.

BibTeX

If you find our work useful in your research, please consider citing:

@misc{cai2026viewpointgeneralizationmultiviewdemonstrations,
      title={Beyond Viewpoint Generalization: What Multi-View Demonstrations Offer and How to Synthesize Them for Robot Manipulation?}, 
      author={Boyang Cai and Qiwei Liang and Jiawei Li and Shihang Weng and Zhaoxin Zhang and Tao Lin and Xiangyu Chen and Wenjie Zhang and Jiaqi Mao and Weisheng Xu and Bin Yang and Jiaming Liang and Junhao Cai and Renjing Xu},
      year={2026},
      eprint={2603.26757},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.26757}, 
}