Multi-view robot policies get an action frame per camera: OVOA-VLA reaches 88.1% average success on LIBERO-Plus
Synopsis
Multi-view vision-language-action models usually receive action supervision only in the robot base frame, so they must implicitly learn the spatial correspondence between each view's observation and base-frame actions. OVOA-VLA instead has each view's action expert predict actions in its own camera frame, so the same physical motion supplies several observation-aligned supervision signals, and two complementary directed attention graphs plus cross-graph consistency feed those predictions into the base policy. On LIBERO-Plus, full and base-only inference reach average success rates of 88.1% and 85.7%; on a real Franka Research 3, base-only inference raises average success from 28.3% to 58.3% across three tasks under spatial distribution shifts.
Interpretation
Action supervision for multi-view vision-language-action models no longer sits only in the robot base frame; every camera view gets an action-prediction task expressed in its own frame. Earlier multi-view models combined observations from several cameras but supplied action supervision only in the base frame, leaving the observation-to-action spatial correspondence to be learned implicitly. On LIBERO-Plus, full inference reaches 88.1% average success and base-only inference 85.7%, both above the reproduced baseline's 77.4% in the table.
The same demonstrated motion is transformed into each camera frame by known coordinate transformations, and each view expert reads only its own view's image and receives robot state in that frame. The method uses a Mixture of Transformers, adding an independently parameterized action expert for each input view alongside the base-frame action expert, with all experts sharing a vision-language backbone. During training each sample is passed forward once under each of two directed attention graphs that share parameters, and a cross-graph consistency loss pulls the two base-frame predictions toward agreement.
The real-robot gains appear under both camera-viewpoint and robot-initialization shifts, not only in simulation. Real experiments use a 7-DoF Franka Research 3 with one fixed third-person camera and one wrist-mounted camera, collecting 50 GELLO teleoperation demonstrations per task. Across six in-distribution tasks, base-only inference averages 77.5% success, 5.0 percentage points above the baseline; under camera-viewpoint shift it rises from 26.7% to 60.0%, and under robot-initialization shift from 30.0% to 56.7%.
Perspective
The demonstrated scope is four simulation benchmarks (LIBERO, LIBERO-Plus, VLABench, MimicGen) and six real tasks on one 7-DoF Franka Research 3; the real-world shift evaluation runs five rollouts per task and configuration, all zero-shot without retraining.
Base-only inference beats full inference under camera shift (78.4% versus 75.4%), while the reverse holds under robot-initialization shift (81.2% versus 88.4%), so each inference mode has its own conditions; full inference raises peak memory and mean latency by roughly 81% and 63%.
