Robot VLA progress is being measured by executable control

Spatial alignment is a main performance lever. Dex-BEV puts visual geometry, proprioception, and output actions into a shared bird’s-eye-view coordinate frame. It reports 97.8% average success on official LIBERO, 76.0% on RoboTwin 2.0 Clean, and 89.9% on modified LIBERO camera and pose settings where listed 2D baselines fall below 10%.

GeoAlign adds RGB-derived geometry features queried by robot state. The gains are clearest on geometry-sensitive tasks: real ALOHA average success is 78.8%, versus 65.0% for the RGB-only baseline, with transparent-bottle success at 75.0% versus 35.0%. 3DThinkVLA takes a lighter deployment route. It trains latent 3D perception and reasoning adapters, then runs on 2D images at inference while reaching 98.7% on LIBERO and 81.0% on LIBERO-PLUS.

World Action Models (WAMs) are being tested by their effect on action choice, latency, and real rollout agreement. GeoSem-WAM trains future RGB, geometry, and semantic prediction heads, then removes those heads at deployment. The reported results are practical: 98.55% average success on LIBERO and 95.4% average success on real Franka tasks, compared with 88.9% for Fast-WAM.

PiL-World uses a policy-in-the-loop setup. A frozen VLA predicts an action chunk, the world model predicts the next multi-view observation, and that generated observation becomes the next policy input. Across three real dual-arm tasks, it cuts the average real-versus-imagined success-rate gap to 12.0%, compared with 63.2% for Ctrl-World, and reports 0.94 Pearson correlation between real and imagined success rates.

Several papers improve policy behavior by changing the action interface or adaptation route while keeping the main VLA mostly fixed. ActionMap replaces single-point action decoding with voxel heatmaps over translation, rotation, and gripper commands. On LIBERO with OpenVLA-OFT, it reaches 97.3% average success versus 89.1% for L1 regression at matched training steps. On real Franka tasks, it succeeds in 20 of 30 trials versus 7 of 30 for the regression head.

WIZARD targets task adaptation. It predicts task-specific LoRA weights from a language instruction and a short demonstration video, then runs a frozen VLA with the generated adapter. On held-out LIBERO-Spatial, it reaches 0.40 average success, compared with 0.19 for the strongest listed multi-task VLA baseline and 0.02 for nearest-neighbor adapter retrieval. The remaining gap to task-specific experts is large, but the evidence shows that adaptation speed and data requirements are now part of the control story.