Multi-view perception is often treated as a safe design choice for robot manipulation: if one camera is useful, more cameras should be better. We revisit that assumption in tabletop bimanual manipulation by comparing policies trained with two large-field-of-view (FoV) wrist cameras against policies that additionally receive a head view, using controlled-variable experiments on a real bimanual platform. Across three tabletop tasks, adding a head view does not consistently improve task completion or execution quality: when the wrist cameras already observe the objects, grippers, and goal regions, wrist-only policies form a strong baseline. Restricting the wrist-camera FoV increases the potential value of a head view, but performance improves most reliably when the added view supplies concentrated, task-relevant information: region-of-interest (ROI) cropping yields more consistent head-view gains than the uncropped image, which is dominated by background and task-irrelevant regions. The data-scaling results further suggest an interaction between FoV and demonstration count: wide observations improve coverage but may require more data to stabilize, and they may reach a higher performance ceiling once enough demonstrations are available. These results suggest that camera design should be framed as a task-relevance and information-density problem rather than as a simple count of input views.
Figures & tables
Figure 1: Overview of the study, organized by its three contributions and its key findings. Top: the guiding question and the experimental setting—a real bimanual platform with two large-FoV wrist cameras and one fixed head camera, evaluated on three tabletop tasks. \scriptsize1⃝ The controlled observation design crosses wrist-camera FoV (Large, D435-like; shown by representative wrist frames) with head-view input (NoHead, RawHead, ROIHead), yielding the controlled configurations G1–G6 evaluated in Sec. 4 . \scriptsize2⃝ Each rollout is scored with a quality-adjusted metric that separates high-quality success from coarse completion, illustrated by two Task B final states. \scriptsize3⃝ RISE-style action-sensitivity diagnostics: in the Task B example the wrist views carry most of the step-level action share ( 0.59+0.26 versus 0.14 for the head view), which contributes low-weight global context. Bottom band: key findings from the two complementary experiments—the Task A data-scaling trend, which suggests that large-FoV input needs more demonstrations but may reach a higher performance ceiling, and the Task B wrist-FoV scan, where the ROIHead-versus-NoHead FinalScore contrast is non-monotonic in wrist FoV ( +21.0 Large, −5.3 FoV120, −7.9 FoV90, +29.2 D435-like).
Figure 2: Task overview and initial layouts. Rows show Tasks A–C, and columns show Init01–Init03. Red annotations specify the task goals, with numbered assignments illustrated in Init01. In Task A, arrows map the four objects to storage positions 1–4. In Task B, the green target basket is circled, and the surrounding baskets to be nested are labeled 1–4. In Task C, arrows map plates 1–3 to rack slots, and boxes mark the rack region.
Symbol
Quantity
Definition
ng
Successful grasps
Times a target object is securely lifted by the gripper.
np
Correct placements
Target objects placed or inserted into the assigned goal region.
ns
Stable placements
Placements where the object retains its intended pose at the target after release.
nr
Retries
Repeated grasp or alignment attempts after a failed attempt.
T
Executed steps
Executed low-level control steps in the rollout.
ng,loc , np,loc
Localized grasps and placements (Task B)
Grasps and placements executed on the correctly localized baskets.
Table 1: Directly observed scoring quantities. Only these quantities are manually recorded for each rollout; all derived scores in this section are computed from them by fixed formulas. Localization-related counts apply to Task B and Task C as indicated.
Figure 3: Camera configuration schematic using representative policy-input frames from the basket-nesting experiments. Top row: left-wrist, right-wrist, RawHead, and ROIHead inputs. Bottom row: representative left-wrist inputs under Large, FoV120, FoV90, and D435-like settings. The panels come from separate rollouts and therefore illustrate view geometry and task-content density rather than a frame-matched crop comparison.
ID
Wrist FoV
Head Type
Input Configuration
Purpose
G1
Large
NoHead
Large Wrist Only
Large-FoV dual wrist baseline.
G2
Large
RawHead
Large Wrist + RawHead
Analyze RawHead under the large-FoV wrist setting.
G3
Large
ROIHead
Large Wrist + ROIHead
Analyze ROIHead under the large-FoV wrist setting.
G4
D435-like
NoHead
Cropped Wrist Only
Simulate restricted wrist vision.
G5
D435-like
RawHead
Cropped Wrist + RawHead
Analyze RawHead under the small-FoV wrist setting.
G6
D435-like
ROIHead
Cropped Wrist + ROIHead
Analyze ROIHead under the small-FoV wrist setting.
Table 3: Six main experiment configurations.
Figure 4: G1–G6 bar plots for the three tasks. Each bar shows FinalScore split into the outcome contribution (solid) and the high-quality bonus (light top segment), following the Score Split column in Table 4 . Bars without a light segment complete rollouts without earning any high-quality bonus.
Figure 5: Summary matrix of mean FinalScore across tasks and configurations. The continuous heatmap uses a muted blue-green palette consistent with the style of Figure 4 ; lighter cells indicate lower FinalScore and darker cells indicate higher FinalScore.
FoV Setting
NoHead
ROIHead
Purpose
Large
F1 / G1
F2 / G3
Large-FoV baseline.
FoV120
F5
F6
Medium-FoV comparison.
FoV90
F7
F8
More restricted FoV comparison.
D435-like
F3 / G4
F4 / G6
Small-FoV comparison.
Table 5: Wrist-FoV experiment configurations.
Figure 6: Decomposition of the head-view contrast across wrist-FoV settings on Task B. (a) The FinalScore contrast Δhead (ROIHead − NoHead) is positive under Large and D435-like but negative at the medium settings. (b) Metric-level changes explain this pattern: the localization rate improves under every FoV setting, whereas the Quality Success Rate flips sign at FoV120 and FoV90, where longer ROIHead trajectories violate the T≤900 high-quality gate (the completion-rate change at FoV120 is zero); see Table 6 for the full per-setting values, including steps. Values are descriptive means over the fixed rollout slots.
Configuration
Wrist FoV
Head Type
Demonstrations
2W-Large
Large
NoHead
50 / 200 / 800
2W-D435Crop
D435-like
NoHead
50 / 200 / 800
3V-D435Crop-ROIHead
D435-like
ROIHead
50 / 200 / 800
Table 7: Data-scaling experiment configurations.
Figure 7: Scaling curves. The x-axis is the number of demonstrations. The y-axis is FinalScore. The curves include 2W-Large, 2W-D435Crop, and 3V-D435Crop-ROIHead, with descriptive bootstrap intervals.
Figure 8: Initial observability in the toy-transfer diagnostic. (a,b) The head view distinguishes whether the toy starts in the left or right bin. (c,d) The initial left- and right-wrist observations do not expose the target because of the bin walls. The panels document the visual information available at the initial decision stage; they are not rollout-success evidence.
RQ
Evidence
What it answers
RQ1
G2 − G1, G3 − G1
How much gain RawHead or ROIHead adds when wrist FoV is large.
RQ2
G5 − G4, G6 − G4; FoV scan (Table 6 ); G1 vs. G6
Whether restricted wrist FoV increases head-view value, and when that potential fails to raise the quality-adjusted score.
RQ3
G3 − G2, G6 − G5
Effect of ROI cropping relative to the raw head view under large and small wrist FoV.
How demonstration count interacts with the visual-input configuration.
Table 8: Research questions and the evidence that serves them. The rows follow the RQ order in the introduction; contrasts are computed from the G1–G6 results in Table 4 .
Task
Comparison
Δ Completion Rate
Δ FinalScore
Δ Quality Success Rate
Δ Steps
Task A
G2 − G1 RawHead
+40 pp
-16.0
-40 pp
+70.0
Task A
G3 − G1 ROIHead
+30 pp
-6.0
-20 pp
+75.0
Task B
G2 − G1 RawHead
-10 pp
-1.1
-10 pp
+67.5
Task B
G3 − G1 ROIHead
+20 pp
+21.0
+30 pp
-47.5
Task C
G2 − G1 RawHead
0 pp
0.0
0 pp
-5.0
Task C
G3 − G1 ROIHead
+20 pp
+12.2
+20 pp
-60.0
Table 9: Head-view contrasts under the Large FoV setting. The comparisons are G2 − G1 and G3 − G1. Values show changes relative to the Large FoV NoHead baseline G1. FinalScore contrasts are computed from unrounded means before one-decimal display.
Figure 9: Head-view contrasts under the Large FoV setting. Horizontal bars show the mean Δ FinalScore for G2 − G1 or G3 − G1 relative to the G1 NoHead baseline. The zero line separates lower and higher mean scores relative to G1.
Figure 10: Representative real-robot final states illustrating the distinction between completion and high-quality completion. Completion-only examples can satisfy coarse task progress while still losing the high-quality bonus because of unstable final pose or poor alignment; long execution and repeated attempts are additional scoring-protocol failure modes summarized in Table A2 .
Figure 11: RISE-style action-sensitivity example for RawHead and ROIHead on Task A. Heatmaps visualize Sv(c) from Eq. F3 ; action-share values report qv from Eq. F4 within each selected step. The maps indicate local action sensitivity, not attention weights or rollout success.
Figure 12: RISE-style action-sensitivity examples across the three tasks. Heatmaps show Sv(c) , and view-level shares are qv within each selected step. Wrist views typically respond near grippers, objects, and contact regions, while the head view provides lower-weight global context. These examples are diagnostic illustrations of view roles and are not success-rate evidence.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B1 : Representative RawHead and ROIHead policy inputs from the basket-nesting experiments. RawHead retains broad scene context with lower task-region concentration, whereas ROIHead concentrates the task-relevant region. The panels come from separate rollouts and illustrate view geometry rather than a frame-matched crop comparison.
Figure C1 : RISE-style action-sensitivity comparison between NoHead and ROIHead under the restricted-wrist-FoV setting. NoHead contains only the two wrist views, while ROIHead adds the head view. In these representative single-step examples, wrist views have the larger qv values, while the head view provides auxiliary spatial cues.
Figure D1 : RISE-style head-view action sensitivity and representative target-box choices in the toy-transfer diagnostic. (a,b) For the two initial toy locations, the dominant high-response region overlaps the occupied bin and shifts with the visible toy location. The overlays visualize Sv(c) and are interpreted spatially; color intensity is not compared quantitatively across panels. (c,d) In the shown wrist-only cases, the policy selects the left bin for both toy locations, yielding a correct box choice when the toy is on the left and an incorrect choice when it is on the right. (e) In the shown case with an added head view, the policy selects the right bin when the toy is on the right. The behavior frames are recorded from the head-camera viewpoint for visualization. These unscored examples do not estimate success rates, prove causal reliance, or establish the cause of the wrist-only left-bin preference.
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.
Boyang Cai, Qiwei Liang, Jiawei Li +11
The Hong Kong University of Science and Technology (Guangzhou) · Shenzhen University · Beijing Jiaotong University +1
Manipulation requires both broad scene awareness and detailed local feedback, yet conventional camera rigs provide them through separate front and wrist cameras. We present Fisheye-VLA, a visual interface that brings these capabilities together using a single passive fisheye. A global view preserves the workspace, while local perspective crops direct detail toward the interaction. The key design question is where this local visual budget should go. We answer it through a controlled re-rendering study, comparing alternative crop directions on the same recorded observations. The study finds that end-effector-centered views capture most of the estimated benefit of a much larger candidate pool, motivating a compact allocation around both hands. Our interface uses calibrated end-effector projection and motion lead to track the crops, while a shared ray encoding preserves their spatial meaning as they move. Integrated with a pretrained VLA, it achieves 84% and 82% success in the two expanded tabletop regions, where some target placements extend beyond the front-camera coverage, and supports shelf and conveyor manipulation. Ablations show that local crops and their viewing directions become more important in the larger workspace regions. The results demonstrate that a single fisheye can support these manipulation tasks without physical wrist cameras.
Ziang Ren, Zike Yan, Raymond Zhang +2
Hong Kong Embodied AI Lab · The Chinese University of Hong Kong, Hong Kong SAR, China · DeepCybo +1
Inspired by human vision, we introduce a framework using active gaze to enable fine-grained bimanual manipulation with only a single stereo camera. EyeRobot 2.0 physically attends to a 3D fixation point in the scene by swiveling two eye viewpoints to center their gaze on it. The resulting images are processed foveally by allocating more visual tokens to the image centers, focusing computation on task-relevant features. Such Active Visual Fixation (AVF) requires carefully coordinated gaze during task execution, which we accomplish hierarchically by first training a low-level gaze servoing policy conditioned on a goal object, then training a target selector which emits fixation goals based on task progress. Both modules are trained with RL on real-world data: the first is trained with a dense geometric reward and the second co-trains with the BC gripper policy which allows it to discover fixation sequences that can resemble a human's fixation sequence while performing the task. EyeRobot 2.0 further takes advantage of fixation by canonicalizing gripper information into a fixation-relative SE(3) frame, which compacts the size of the action distribution to learn. We collect teleoperation data for 7 real-world and 6 simulated tasks, and conduct over 1000 physical and 1800 simulated robot trials comparing EyeRobot 2.0 against passive stereo and ego + wrist camera policies trained on the same data. Removing wrist cameras is costly for standard policies: with only passive stereo, real-world success drops from 52% to 27%. EyeRobot 2.0 closes this gap with only stereo, outperforming passive stereo by 40% in real and 20% in sim. It matches ego + wrist policies when their wrist views are clear (69% vs. 64%), and more than doubles their success when grasped objects occlude the wrist cameras (48% vs. 22%)