Multi-Camera View Scaling for Data-Efficient Robot Imitation Learning

Yichen Xie; Yixiao Wang; Shuqi Zhao; Cheng-En Wu; Masayoshi Tomizuka; Jianwen Xie; Hao-Shu Fang

cs.ROApr 1, 2026

Multi-Camera View Scaling for Data-Efficient Robot Imitation Learning

Authors: Yichen Xie, Yixiao Wang, Shuqi Zhao, Cheng-En Wu, Masayoshi Tomizuka, Jianwen Xie, Hao-Shu Fang

Organizations: UC Berkeley · Lambda, Inc · MIT

Abstract

The generalization ability of imitation learning policies for robotic manipulation is fundamentally constrained by the diversity of expert demonstrations, while collecting demonstrations across varied environments is costly and difficult in practice. In this paper, we propose a practical framework that exploits inherent scene diversity without additional human effort by scaling camera views during demonstration collection. Instead of acquiring more trajectories, multiple synchronized camera perspectives are used to generate pseudo-demonstrations from each expert trajectory, which enriches the training distribution and improves viewpoint invariance in visual representations. We analyze how different action spaces interact with view scaling and show that camera-space representations further enhance diversity. In addition, we introduce a multiview action aggregation method that allows single-view policies to benefit from multiple cameras during deployment. Extensive experiments in simulation and real-world manipulation tasks demonstrate significant gains in data efficiency and generalization compared to single-view baselines. Our results suggest that scaling camera views provides a practical and scalable solution for imitation learning, which requires minimal additional hardware setup and integrates seamlessly with existing imitation learning algorithms. The website of our project is https://yichen928.github.io/robot_multiview.

Figures & tables

Fig. 1 : Camera View Scaling-Up. Each expert demonstration can be converted to multiple pseudo-demonstrations by exploiting multiple camera perspective views.

Fig. 2 : Action Space for Pseudo-Demonstrations. We consider three potential action spaces: base , camera , and EEF .

Training View	Space	Square			Can			Lift
Training View	Space	$N=10$	$N=25$	$N=50$	$N=10$	$N=25$	$N=50$	$N=10$
$Cam_{F}$	Base	0.14	0.26	0.42	0.18	0.54	0.72	0.69
$Cam_{F},Cam_{FL},Cam_{FR}$	Base	0.16 ( 14% $\uparrow$ )	0.31 ( 19% $\uparrow$ )	0.47 ( 12% $\uparrow$ )	0.27 ( 50% $\uparrow$ )	0.63 ( 17% $\uparrow$ )	0.78 ( 8% $\uparrow$ )	0.79 ( 14% $\uparrow$ )
$Cam_{F},Cam_{FL},Cam_{FR}$	Camera	0.17 ( 21% $\uparrow$ )	0.29 ( 12% $\uparrow$ )	0.47 ( 12% $\uparrow$ )	0.30 ( 67% $\uparrow$ )	0.68 ( 26% $\uparrow$ )	0.83 ( 15% $\uparrow$ )	0.76 ( 10% $\uparrow$ )
$Cam_{F},Cam_{FL},Cam_{FR}$ $Cam_{FU},Cam_{FD}$	Base	0.18 ( 29% $\uparrow$ )	0.42 ( 62% $\uparrow$ )	0.55 ( 31% $\uparrow$ )	0.32 ( 78% $\uparrow$ )	0.65 ( 20% $\uparrow$ )	0.80 ( 11% $\uparrow$ )	0.80 ( 16% $\uparrow$ )
$Cam_{F},Cam_{FL},Cam_{FR}$ $Cam_{FU},Cam_{FD}$	Camera	0.18 ( 29% $\uparrow$ )	0.39 ( 50% $\uparrow$ )	0.58 ( 38% $\uparrow$ )	0.37 ( 105% $\uparrow$ )	0.68 ( 26% $\uparrow$ )	0.85 ( 18% $\uparrow$ )	0.85 ( 23% $\uparrow$ )

TABLE I : Camera View Scaling-Up Results in Simulation Environments.

Panel of Figure 4 — Fig. 3 : Visualization of Attention Scores. When trained with randomly picked camera views, the model learns to focus on the most important regions to finish the task. Otherwise, the attention is distracted to unrelated regions.

Training View	Space	Square
Training View	Space	$N=10$	$N=25$	$N=50$
$Cam_{F}$	Base	0.14	0.26	0.42
$Cam_{F}$	EEF	0.06	0.25	0.32
$Cam_{F},Cam_{FL},Cam_{FR}$ $Cam_{FU},Cam_{FD}$	Base	0.18	0.42	0.55
$Cam_{F},Cam_{FL},Cam_{FR}$ $Cam_{FU},Cam_{FD}$	EEF	0.09	0.27	0.41

TABLE II : Effect of Action Space.

Training View	Square
Training View	$N=10$	$N=25$	$N=50$
$Cam_{F}$	0.14	0.26	0.42
$Cam_{F},Cam_{FL}^{30},Cam_{FR}^{30}$	0.18	0.26	0.44
$Cam_{F},Cam_{T},Cam_{S}$	0.17	0.24	0.40
$Cam_{F},Cam_{FL},Cam_{FR}$	0.17	0.29	0.47

TABLE III : Selection of Camera Views.

Training View	Inference View	Can
Training View	Inference View	$N=10$	$N=25$
Five Views	$Cam_{F}$	0.37	0.68
	$Cam_{FL}$	0.30	0.66
	$Cam_{FU}$	0.37	0.77
Single View	$Cam_{F}$	0.18	0.54
	$Cam_{FL}$	0.00	0.07

TABLE IV : Camera View Generalization in the Inference.

Training View	Inference View	Square			Can
Training View	Inference View	$N=10$	$N=25$	$N=50$	$N=10$	$N=25$	$N=50$
$Cam_{F}$	$Cam_{F}$	0.14	0.26	0.42	0.18	0.54	0.72
$Cam_{F},Cam_{FL},Cam_{FR}$	$Cam_{F}$	0.17	0.29	0.47	0.30	0.68	0.83
$Cam_{F},Cam_{FL},Cam_{FR}$	$Cam_{F},Cam_{FL},Cam_{FR}$	0.18	0.33	0.51	0.39	0.73	0.85

TABLE V : Multiview Composition with Single-View Policy.

Fig. 4 : Real-World Robot Setup. We use a FANUC CRX-10iA robotic arm equipped with a parallel-jaw gripper with silicone soft fingertips and two camera views.

Training View	Composition	$N=25$	$N=50$
Single	✗	0.45	0.70
Both	✗	0.50	0.85
Both	✓	0.75	0.85

TABLE VI : Real-World Experiment Results. We consider whether to train with pseudo-demonstration and infer with multiview aggregation.

Explore similar work

Beyond Viewpoint Generalization: What Multi-View Demonstrations Offer and How to Synthesize Them for Robot Manipulation?
Mar 23, 2026Boyang Cai, Qiwei Liang, Jiawei Li +11Multi-View Learning Generalization in Robotic Manipulation
The Hong Kong University of Science and Technology (Guangzhou) · Shenzhen University · Beijing Jiaotong University +1
Data Scaling Laws in Imitation Learning for Robotic Manipulation
Oct 24, 2024Fanqi Lin, Yingdong Hu, Pingyue Sheng +3Generalization in Robotic Manipulation Robotic Manipulation
Tsinghua University · Shanghai Qi Zhi Institute · Shanghai Artificial Intelligence Laboratory
RESTA: Time-Scaling Augmentation for Data-Efficient Robot Imitation Learning
Oct 13, 2026Yunpeng Jiang, Jianshu Hu, Han Fang +3Data-Efficient Robot Learning Robot Imitation Learning
Global College, Shanghai Jiao Tong University, Shanghai, China · Shanghai Institute of Physical Intelligence and Robotics (PAIR), Shanghai, China · Duke Kunshan University, Jiangsu, China

Mar 23, 2026J/K move · Enter open · S save

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

Boyang Cai, Qiwei Liang, Jiawei Li, Shihang Weng+10

The Hong Kong University of Science and Technology (Guangzhou) · Shenzhen University · Beijing Jiaotong University+1

Multi-View Learning Generalization in Robotic Manipulation Robotic Manipulation Multi-View Video Generation

Open paper AskPDF arXiv

Does multi-view demonstration truly improve robot manipulation, or merely enhance cross-view robustness? We present a systematic study quantifying the performance gains, scaling behavior, and underlying mechanisms of multi-view data for robot manipulation. Controlled experiments show that, under both fixed and randomized backgrounds, multi-view demonstrations consistently improve single-view policy success and generalization. Performance varies non-monotonically with view coverage, revealing effective regimes rather than a simple "more is better" trend. Notably, multi-view data breaks the scaling limitation of single-view datasets and continues to raise performance ceilings after saturation. Mechanistic analysis shows that multi-view learning promotes manipulation-relevant visual representations, better aligns the action head with the learned feature distribution, and reduces overfitting. Motivated by the importance of multi-view data and its scarcity in large-scale robotic datasets, as well as the difficulty of collecting additional viewpoints in real world settings, we propose RoboNVS, a geometry-aware self-supervised framework that synthesizes novel-view videos from monocular inputs. The generated data consistently improves downstream policies in both simulation and real-world environments.

Abstract

Explore similar work

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

Data Scaling Laws in Imitation Learning for Robotic Manipulation

RESTA: Time-Scaling Augmentation for Data-Efficient Robot Imitation Learning