WEVO: World Model-Enabled VLA Optimization for Whole-Body Loco-Manipulation (Nov 20). The WEVO preprint combines simulation data with physical rollouts to train a video world model for improving a humanoid policy. Teams can use the learned environment to explore policy changes before testing them on hardware.1
Readiness (Research prototype). The starting policy collected 20 autonomous robot rollouts per task to align the world model with physical observations.
Pros. Mean success on three physical tasks rose from 23.3% to 75.0% after world-model reinforcement learning in the authors’ tests.
Cons. The approach still needs target-robot rollouts, and evidence is limited to the three tasks.
Digital Twin Calibration
Fixing the Twin, Not the Detector: Sim-to-Real Roadside LiDAR Detection from Digital-Twin Data (Nov 20). This preprint aligns a roadside lidar twin to real coordinates, point density, and annotation conventions before training a detector on synthetic labels. Calibration offers a concrete way to improve transfer when real labeled data from fixed sensors is scarce.2
Readiness (Research prototype). The authors compared synthetic-to-real transfer on two UrbanTwin challenge tracks and used unlabeled target frames to match point density.
Pros. The approach placed first on the LUMPI track with a 0.5541 combined challenge score.
Cons. The workflow relies on access to unlabeled target-sensor data, and some validation gains did not carry through to test.