Source-linked AI summary
WARP-VLA: Wrist-Camera Adaptation for View-Robust Policy Execution in Vision-Language-Action Models
Junmyeong Lee, Dongmin Shin, Min-Gyu Park, Wooseok Jeon, Inho Chang, Hae-Gon Jeon
TL;DR
Camera-configuration changes make VLA deployment difficult, especially for wrist cameras where small shifts alter interaction cues. WARP-VLA uses view-specific experts and feature-conditioned routing without camera calibration; under wrist-camera perturbations, π0.5's average success rate rises from 39.2% to 78.3%.
Problem
VLA performance is sensitive to camera changes, but reproducing the training configuration is difficult and wrist-view shifts can alter fine-grained interaction cues.
Source: §I · p. 1Method
WARP-VLA adapts wrist-camera features with view-specific experts and feature-conditioned routing, without requiring camera calibration.
Source: §I · p. 2Results
π0.5's average success rate under wrist-camera perturbations rises from 39.2% to 78.3% with WARP-VLA.
Source: §I · p. 2Takeaways & Limitations
The paper releases a wrist-view robustness benchmark and a plug-and-play implementation.
Source: §I · p. 2
Abstract
from arXiv · showhide
Despite recent advances in Vision-Language-Action models (VLAs) for robotic manipulation, their performance remains sensitive to changes in camera configuration. The problem becomes more evident in cross-setup deployment, as reproducing the exact camera pose used for training is nearly impossible. Unlike fixed external views, wrist views are more challenging because the camera moves with the robot, causing even small mounting variations to alter fine-grained geometric cues. To address this, we propose WARP-VLA, a camera-view robust VLA for diverse wrist camera configurations. WARP-VLA adopts a Mixture-of-Experts (MoE) architecture where individual experts learn view-specific feature transformations, and a router combines them based on implicit view information. This allows the policy to be deployed without requiring camera extrinsic parameters as additional input. Through experiments on the LIBERO benchmark, WARP-VLA improves the average success rate of pi-0.5 from 39.2% to 78.3% under wrist-view perturbations. The real-robot experiments further show that the feature-level adaptation learned in simulation successfully transfers to diverse deployment settings. To facilitate reproducibility and future research, we release our wrist viewpoint robustness benchmark and a plug-and-play implementation.
I. INTRODUCTION
Wrist-camera changes can disrupt VLA deployment by altering fine-grained interaction cues, while reproducing the training camera configuration is difficult. WARP-VLA adapts wrist-view features without camera calibration; experiments report improved robustness, including a higher π0.5 success rate under wrist-view perturbations.
I. INTRODUCTION
- Wrist-camera shifts can alter gripper–object spatial relations, while reproducing the original configuration is difficult across mounting arrangements.More broadly, camera configurations are often undocumented, and mounting tolerances or robot vibrations can introduce viewpoint changes.Source: §I · pp. 1–2
- Prior camera-robustness methods often rely on camera geometry or synthesized target views, which require calibration or camera-specific models.For wrist cameras, inaccurate calibration can compromise interpretation of close-range interactions.Source: §I · p. 2
- WARP-VLA uses view-specific feature transformations and feature-conditioned routing to adapt wrist views without camera calibration.The approach is presented as a plug-and-play module for VLA policies.Source: §I · p. 2
- π0.5's average success rate under wrist-camera perturbations rises from 39.2% to 78.3% with WARP-VLA.The paper also reports successful real-robot grasping after camera position changes, where the baseline struggles to locate the object.Source: §I · p. 2
- The evaluation adapts LIBERO benchmarks into a wrist-view protocol and reports robustness across simulation and real-robot settings.The contributions include a view-augmented LIBERO dataset with diverse wrist-camera configurations.Source: §I · p. 2
II. RELATED WORK
Prior VLA work addresses viewpoint variation through view augmentation, camera conditioning, view synthesis, and geometric representations. The paper distinguishes its objective as adapting perturbed wrist features to a frozen VLA's canonical representations.
A. View-Invariant VLAs
- View-invariant VLA approaches expand training views, add geometric or camera cues, convert camera-frame actions, or synthesize test views, but coverage and domain adaptation remain concerns.AnyCamVLA avoids policy retraining but its novel-view-synthesis model still requires simulation-domain fine-tuning on multi-view images.Source: A. View-Invariant VLAs · p. 2
B. Geometry-Aware VLAs
- Geometry-aware VLAs add depth or point-cloud cues, integrate geometric priors into policy components, or align features with geometry during training.Adapting perturbed wrist features to a frozen VLA's canonical representations remains a distinct objective.Source: B. Geometry-Aware VLAs · pp. 2–3
A. Problem Definition
WARP-VLA adapts perturbed wrist-camera features toward the canonical feature space without explicit camera parameters at inference. Its MoE adapter combines learned view-specific transformations, using a pseudo-canonical anchor while retaining the frozen vision encoder and action policy.
A. Problem Definition
- WARP-VLA maps perturbed-view features toward canonical features, training the adapter to recover canonical representations for the action policy.The predicted canonical features replace the perturbed tokens entering the policy.Source: A. Problem Definition · p. 3
- The pseudo-canonical anchor averages wrist-view features across frames in the base policy’s canonical-view fine-tuning dataset.Averaging across scenes and tasks provides a stable reference associated with the canonical camera configuration.Source: A. Problem Definition · p. 3
B. Canonical-View Adaptation with a Mixture of Experts
- A finite set of learned experts models mappings across the continuous wrist-camera configuration space, with a feature-conditioned router combining their outputs.The adapter is inserted between the image encoder and action policy, preserving the token shape expected by the policy backbone.Source: B. Canonical-View Adaptation with a… · p. 3
- Paired canonical features supervise the MoE adapter while the vision encoder and action policy remain frozen.The adapter trained on LIBERO is reused in real-robot experiments without further adapter training.Source: Figure 2 · p. 3, B. Canonical-View Adaptation with a… · p. 3
C. Anchor-Guided Routing with Spatial Token Statistics
WARP-VLA routes wrist-view features among experts using statistics of the perturbed features and a canonical anchor. Separate routing groups represent translation and tilt variations, with feature statistics supplying cues for view deviation.
C. Anchor-Guided Routing with Spatial Token Statistics
- The router predicts expert contributions from how the observed wrist view deviates from the canonical representation.Its input is constructed from the perturbed-view features and canonical anchor, and randomized anchor mixing during training discourages overfitting to one canonical view.Source: C. Anchor-Guided Routing with Spatial… · p. 3
- The router representation combines channel-wise activation and spatial-token statistics from the perturbed-anchor residual and anchor.Spatial statistics provide cues for distinguishing perturbation axes, while comparing residual and anchor features reduces reliance on scene- and task-specific visual content.Source: C. Anchor-Guided Routing with Spatial… · pp. 3–4
- Figure 3 depicts 13 radial-angle target regions and router-selected radial and angular groups, with point locations showing ground-truth offsets and colors showing routing predictions.The router uses separate distributions for translation and tilt expert groups.Source: Figure 3 · p. 4, C. Anchor-Guided Routing with Spatial… · p. 4
- WARP-VLA uses 15 experts: 13 cover translation regions in radial-distance and azimuth space, while two cover tilt; routing is normalized independently by group.The factorized routing represents combinations of translation and tilt variations without competition among unrelated view factors.Source: C. Anchor-Guided Routing with Spatial… · p. 4
D. View-Specific Experts with Low-Rank Factorization
Each view-specific expert transforms perturbed wrist features toward the canonical feature space through cross-attention to a pseudo-canonical anchor. Experts share a transformation and add compact low-rank residuals that specialize them to different view regions.
D. View-Specific Experts with Low-Rank Factorization
- Each expert takes perturbed wrist tokens as queries and the pseudo-canonical anchor as keys and values to produce canonical-view features.The final adapted representation combines expert outputs using routing weights.Source: D. View-Specific Experts with Low-Rank… · p. 4
- Experts share a cross-attention Transformer transformation and use expert-specific low-rank residuals to specialize for different view regions.The shared component captures common adaptation across views, while residuals capture view-dependent corrections without duplicating a full Transformer per expert.Source: D. View-Specific Experts with Low-Rank… · p. 4
- Each expert’s shared projection matrix corresponds to an attention or feed-forward projection in a Transformer block.The expert-specific low-rank factors add a small number of parameters while maintaining dedicated view transformations.Source: D. View-Specific Experts with Low-Rank… · p. 4
- All experts begin with the same canonical-recovery function and gradually specialize during training.The expert-specific residual factors are initialized so the experts initially share the same cross-attention transformation.Source: D. View-Specific Experts with Low-Rank… · p. 4
E. Training Objective
WARP-VLA trains view-specific experts to reconstruct canonical-view features and trains a router to select experts using estimated responsibilities. Training-time pose supervision is not needed at inference, when routing uses only the input wrist-view features.
E. Training Objective
- Expert reconstruction and routing supervision losses jointly encourage expert specialization and view-aware routing.The expert loss measures reconstruction independently for each expert, while the router is trained to match estimated expert responsibilities.Source: E. Training Objective · pp. 4–5
- Each expert is supervised with responsibility-weighted reconstruction to recover canonical-view features and specialize within its camera-configuration region.Independent expert-level reconstruction preserves spatial correspondence for feature adaptation.Source: E. Training Objective · p. 5
- Responsibilities combine routing probability, reconstruction quality, and a training-time pose prior, with overlapping pose bins sharing responsibility near boundaries.The teacher pose prior and canonical targets are used only during training; at inference, the router predicts weights from residual wrist-view features without camera parameters or policy fine-tuning.Source: E. Training Objective · p. 5
A. Implementation Details
The implementation trains WARP-VLA on view-augmented LIBERO demonstrations and uses a shared expert transformation with lightweight routed experts. The paper also illustrates the real-robot workspace and camera observations.
A. Implementation Details
- 2,000 LIBERO demonstrations across 40 tasks and four suites are re-rendered under ten perturbed wrist-camera configurations each.Offsets are sampled within a semicircular translation region of r ≤ 6 cm and θ ∈ [0°, 180°], and tilt range τ ∈ [0°, 10°]; other settings remain fixed.Source: A. Implementation Details · p. 5
- WARP-VLA uses a shared eight-block ViT transformation with 15 routed experts and rank-16 low-rank parameterization.Training runs for 10,000 steps with Adam, batch size 4,096, and learning rate 3 × 10−4, on eight GPUs for approximately 30 hours.Source: A. Implementation Details · p. 5
- The real-robot setup figure shows the manipulation workspace and example observations from external and wrist cameras.The figure provides visual context for the two camera viewpoints.Source: Figure 4 · p. 5
B. Simulation Experiments
WARP-VLA improves wrist-view robustness in simulation, outperforming camera-aware baselines under perturbations and transferring to LIBERO-Plus without further fine-tuning. The evaluation tests canonical and perturbed camera configurations across both benchmarks.
B. Simulation Experiments
- The evaluation covers canonical, small, medium, and large wrist-camera conditions, with 14 configurations at each nonzero perturbation level.LIBERO uses 560 episodes per perturbation level; LIBERO-Plus cycles through the 14 configurations over 10K episodes per level.Source: B. Simulation Experiments · p. 5
- π0.5 averages 39.2% success on LIBERO, while GAM falls from 98.4% canonical success to 33.4% under large perturbations.The table summarizes success rates under wrist-view perturbations on LIBERO and LIBERO-Plus.Source: B. Simulation Experiments · p. 6, Table I · p. 6
- 6.0 and 16.0 percentage points: WARP-VLA’s average success rate across perturbed levels exceeds AnyCamVLA and KYC, respectively.WARP-VLA also achieves a higher success rate than AnyCamVLA at the small perturbation level; the reported comparison is attributed to feature-space adaptation rather than canonical-view synthesis.Source: B. Simulation Experiments · p. 6, Table I · p. 6
- 56.5%: WARP-VLA with π0.5 achieves the highest LIBERO-Plus average success rate, versus 54.7% for AnyCamVLA and 30.1% for KYC.The LIBERO-trained adapter transfers without additional fine-tuning; it improves average success rates of π0.5 and GR00T N1.7 by 27.7 and 19.1 percentage points, respectively.Source: B. Simulation Experiments · p. 6, Table I · p. 6
C. Real-Robot Experiments
On a real robot, WARP-VLA outperforms π0.5 and AnyCamVLA under wrist-camera perturbations. The experiments use fabricated mounts at multiple perturbation levels, and qualitative comparisons show improved gripper–object alignment.
C. Real-Robot Experiments
- The real-robot evaluation uses one canonical and nine perturbed mounts, with five trials per task across 10 tasks for each camera configuration.Perturbations include x- or y-axis translations of 2, 4, or 6 cm and tilts of 3°, 6°, or 9°; baseline policies are fine-tuned on 50 demonstrations per task.Source: C. Real-Robot Experiments · p. 6
- 48.7%: WARP-VLA’s average success rate under perturbed camera configurations, improving on π0.5 by 12.0 and AnyCamVLA by 8.0 percentage points.WARP-VLA achieves 80% under the canonical setting; the success-rate figure reports results across perturbation levels.Source: C. Real-Robot Experiments · p. 7, Figure 6 · p. 7
- In qualitative trials, π0.5 and AnyCamVLA misplace grasps, while WARP-VLA centers the object between the gripper fingers and completes the cup-and-ball task.The figure’s close-ups show the boxed regions where baseline gripper–object misalignment is visible.Source: C. Real-Robot Experiments · p. 7, Figure 5 · p. 7
D. Ablation Studies
Ablations indicate that robust wrist-view adaptation benefits from view-specific experts and learned input-dependent routing. The conclusion bounds generalization by the training camera distribution and available visual evidence.
D. Ablation Studies
- A single shared adapter improves on the unadapted VLA but is less robust than WARP-VLA across perturbed views.The authors attribute this gap to heterogeneous feature changes caused by wrist-camera displacement and tilt.Source: D. Ablation Studies · p. 7
- Uniformly combining view-specific experts yields lower robustness than learned routing, despite using the same expert capacity.The results indicate that MoE benefits from adapting expert contributions to the observed wrist view.Source: D. Ablation Studies · p. 7
- Router-selected assignments form concentric bands and angular sectors that align with the defined camera-offset regions, despite some boundary overlap.Offsets are not router inputs; the alignment indicates that wrist image tokens contain cues for view-dependent routing.Source: D. Ablation Studies · p. 8
V. CONCLUSION
- WARP-VLA improves robustness across LIBERO and LIBERO-Plus, while generalization remains constrained by the training camera distribution and available visual evidence.It cannot fully compensate when task-relevant regions are completely occluded or outside the field of view.Source: §V · p. 8