Diffusion Policies Power Up, Spatial Reasoning Gets Benchmarked, and Humanoids Learn Whole-Body Control

Research

DiffusionVLA Brings Diffusion-Based Action Generation to Vision-Language-Action Models

This merges diffusion model expressiveness with VLA architectures, which is a big deal because diffusion policies handle multimodal action distributions way better than simple regression heads. If this scales, it could be the default action head for language-conditioned robot control.

Whole-Body Locomotion Policy for Humanoids Learns Unified Control from Scratch

Training a single policy for full-body humanoid locomotion rather than separate upper/lower body controllers has been a pain point for anyone deploying humanoids. This approach getting end-to-end whole-body control working is worth watching if you're building on humanoid platforms.

Spatially-Aware Manipulation of Articulated Objects Like Doors and Drawers

Articulated objects are everywhere in homes and factories but notoriously hard for robots because they require understanding joints, constraints, and force dynamics simultaneously. A method that bakes spatial structure into the manipulation policy could unlock a lot of real-world utility.

Navigation World Models Let Robots Plan Paths by Imagining Future Observations

Instead of just reacting to current sensor input, this approach has robots mentally simulate what they'll see at different locations before committing to a path. World models for navigation are heating up because they dramatically reduce sample inefficiency in novel environments.

Benchmarks

ManipBench: New Benchmark Exposes Where Vision-Language Models Fail on Manipulation

VLMs look great on static benchmarks but fall apart when you ask them to reason about spatial relationships for actual grasping. This benchmark systematically catalogs those failures, which should help researchers stop overfitting to easy object-recognition tasks.

Analysis

Rethinking Spatial Intelligence for Embodied Agents

Most robot perception still treats space as a flat 2D projection problem, but real manipulation and navigation demands true 3D spatial reasoning. This work challenges assumptions baked into current architectures and points toward more geometrically grounded representations.