Humanoid Manipulation Matures, Vision-Language Control Expands, and Robot Navigation Gets Smarter

Whole-Body Control for Humanoids Using Hierarchical Reinforcement Learning

This paper tackles one of the hardest problems in humanoid robotics: getting the whole body to coordinate during manipulation tasks. The hierarchical RL approach lets robots decompose complex tasks without needing hand-coded controllers for every subtask.

Vision-Language-Action Models Can Now Generalize Across Robot Morphologies

Most VLA models are tied to specific hardware, which kills portability. This work shows that with the right abstraction layer, a single VLA model can control robots with different arm configurations, which is a big deal for anyone building across hardware platforms.

Zero-Shot Navigation in Unstructured Environments via Semantic Scene Graphs

Getting robots to navigate spaces they have never seen before is still unsolved at scale. This approach builds semantic scene graphs on the fly and uses them to guide navigation without any environment-specific training, which could meaningfully reduce deployment overhead.

Sim-to-Real Transfer for Deformable Object Manipulation Using Adaptive Domain Randomization

Deformable objects like cables, cloth, and food items are notoriously hard to simulate accurately, making sim-to-real transfer brittle. This adaptive randomization method narrows that gap and is directly applicable to assembly and food-handling automation.

Efficient Robot Learning from Sparse Human Feedback Using Active Query Selection

Most imitation learning setups require dense human demonstrations, which is expensive and slow. This active query method figures out which specific moments actually need human input and ignores the rest, cutting annotation cost significantly.