Robot VLA work is concentrating on OOD adaptation with preserved priors and better action structure

HarmoWAM treats manipulation as a phase-dependent control problem. A video world model predicts future frames, while two action experts handle different parts of the task: a reactive expert for transit and a predictive expert for precise interaction. A learned gate chooses the expert during inference.

The paper reports tests on six real-world tasks with background, position, and object variation. In out-of-distribution settings, HarmoWAM claims average gains of 33 percentage points over prior VLA models and 29 points over prior World Action Models. The key detail is the separation between reaching and contact control, which the paper measures directly in its motivation study.

PriorVLA and UniSteer both keep pretrained policy behavior visible during adaptation. PriorVLA freezes a prior expert and trains a separate adaptation expert, updating about one quarter as many parameters as full fine-tuning. It reports 53% Hard OOD success on RoboTwin 2.0 and 57% OOD success in real-world tests across eight tasks and two embodiments.

UniSteer uses a different route for the same deployment pressure. It freezes the flow-matching decoder and trains a small noise actor. Human corrections are mapped back into noise targets, so corrective actions and reinforcement learning update the same actor. Across four real-world tasks, reported average success reaches 90% after 66 minutes of adaptation, compared with 55% for DSRL and 60% for DAgger.

ALAM targets the cost of action-labeled robot data. It learns latent transitions from action-free videos, then regularizes them with composition and reversal constraints. During VLA training, those latent transitions become auxiliary targets while the executed output remains the robot action stream.

The reported gains are large on standard manipulation suites. On MetaWorld MT50, π0 with ALAM reaches 85.0% average success, compared with 47.9% for π0. On LIBERO, it reports 98.1% average success, above the π0 baseline at 94.1%. The result supports a practical reading: unlabeled video is useful when the latent action space carries reusable transition structure.

VEGA focuses on a narrower but important failure source: visual encoders trained mainly on 2D images can miss depth, height, and relative position cues. The method aligns OpenVLA-OFT visual features to a frozen 3D-aware DINOv2-FiT3D teacher during training. The teacher and projector are removed at inference, so runtime stays the same as the base model.

On RoboTwin 2.0, VEGA reports 67.5% average success on Easy and 30.7% on Hard across six bimanual tasks. It improves OpenVLA-OFT by 11.5 points on Easy and 8.0 points on Hard, with task-level gains on Move Card Away, Click Bell, and Place Shoe.