Geometry enters the control loop for robot policies

Vision-Language-Action (VLA) policies are being built around predicted 3D structure and future contact. STARRY jointly denoises future spatial-temporal latents and actions, then uses predicted depth and end-effector geometry to bias attention toward handles, openings, contact surfaces, and nearby obstacles. It reports 93.82% clean and 93.30% randomized success on 50 RoboTwin 2.0 bimanual tasks, with real ARX R5 experiments averaging 70.8% success across three tasks.

X-WAM takes a broader world-action route. It predicts multi-view RGB-D video, robot states, and 32 future actions in one diffusion model. Its asynchronous denoising schedule lets actions decode with fewer steps than video, which matters for closed-loop control. The paper reports 79.2% average success on RoboCasa and 90.7% on RoboTwin 2.0 Randomized, while also evaluating visual and geometric prediction metrics.

The 3D generation survey sets a practical bar for embodied AI assets. Generated objects and scenes need valid geometry, physical parameters, executable kinematics, and simulator-compatible files. That means joints, mass, friction, material behavior, collision geometry, and formats such as URDF, MJCF, and USD matter as much as appearance.

The survey organizes 3D generation into object asset generation, interactive simulation environments, and sim-to-real support. It also compares major simulation platforms including MuJoCo, Isaac Sim, Habitat, AI2-THOR, OmniGibson, PyBullet, ManiSkill3, and Genesis. The open problems are concrete: limited physical annotations, weak agreement between visual quality and physical validity, fragmented evaluation, and sim-to-real gaps.

Walk With Me extends language-conditioned robotics into long outdoor routes. The system maps an abstract request to a concrete destination using GPS context and public map points of interest, then queries a walking-route API and turns the route into waypoints. A Vision-Language Model (VLM) handles destination grounding and safety reasoning, while a low-level VLA policy predicts local motion.

The evidence is mainly a real-world system demonstration. The evaluation covers 20 outdoor trials across last-mile delivery and blind guidance scenarios on an Athena 2.0 Pro robot. The paper claims kilometer-scale operation, but the available summary gives no success rate, collision rate, path-length distribution, completion time, or full-system baseline comparison.