One View, One Action: Spatially Generalizable VLAs through Observation-Aligned Experts and Cross Graph Consistency

TL;DR OVOA-VLA fine-tunes π0.5 with extra action experts that learn each demonstration in every camera's own frame, while the robot still executes base-frame actions. On a real Franka arm under unseen camera views and shifted start poses, success rises from 28.3% to 58.3% at inference cost comparable to π0.5.
What they do
Multi-view VLA policies are supervised only with actions in the robot base frame, so they must learn implicitly how each camera view maps to base-frame motion, and that mapping can become tied to the camera placement and start poses seen in training. OVOA-VLA also expresses every demonstrated action in each camera’s frame and uses these as extra training targets.
How it works
- Per-view targets. Using calibrated extrinsics, the demonstrated end-effector state and action chunk are transformed into each camera’s frame, with no extra annotation.
- Action experts. Starting from π0.5, a mixture of transformers adds one action expert per camera beside the base-frame expert. Each view expert sees only its own camera’s tokens and the state in that frame.
- Two attention graphs. In one, view experts attend to the base expert; in the other, the base expert attends to the view experts. Each sample is trained under both, with a consistency loss between the two base-frame predictions.
- Inference. Full inference uses every expert; base-only inference drops the view experts and keeps FLOPs, memory and latency comparable to π0.5.
Results
- Simulation: 98.7% average on LIBERO (full). On LIBERO-Plus, 88.1% full and 85.7% base-only, 10.7 points above π0.5 for full; camera perturbations rise from 66.2% to 75.4% and robot-initialization perturbations from 67.2% to 88.4%.
- Real Franka FR3, six tasks with 50 GELLO demonstrations each, 20 rollouts per task, base-only: 77.5% vs 72.5% for π0.5; on the three long-horizon tasks, 73.3% vs 65.0%.
- Real spatial shifts on three tasks: unseen third-person cameras 26.7% to 60.0%; start pose offset by 10 cm and 45° of yaw 30.0% to 56.7%.
Limitations
Every camera needs calibrated extrinsics, including hand-eye calibration of the wrist camera. Full inference raises peak memory by about 81% and latency by about 63%. The real-robot shift results rest on five rollouts per task and configuration, and in distribution OVOA-VLA ties π0.5 on three of the six tasks.
Figures




Abstract
Multi-view vision-language-action (VLA) models still face challenges in spatial generalization, particularly under changes in camera viewpoints and robot initial states. These models typically receive action supervision only in the robot base frame, requiring them to implicitly learn the spatial correspondence between observations from each view and base-frame actions. We propose One View, One Action (OVOA-VLA), which learns to predict actions in each view's camera frame, allowing the same physical motion to provide diverse, observation-aligned supervision across views. OVOA-VLA jointly trains a base-frame expert and view-specific action experts, whose robot-state inputs and action targets are expressed in their respective camera frames. Two complementary directed attention graphs model both directions of interaction between the base and view-specific experts, jointly leveraging view-specific action supervision to improve base-frame prediction. Cross Graph Consistency further encourages agreement between the two graphs' base-frame predictions for the same sample. With the same multi-view inputs, the model supports both full and base-only inference. Experiments across multiple simulation benchmarks demonstrate improved manipulation performance and robustness to camera-viewpoint and robot-initialization perturbations. Full and base-only inference achieve average success rates of 88.1% and 85.7% on LIBERO-Plus, respectively. In real-world evaluations, base-only inference raises the average success rates from $π_{0.5}$'s 28.3% to 58.3% across three tasks under spatial distribution shifts.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
