World Models, Unpaired Video Manipulation, and Federated Fleets on the Rise

Object-Centric World Models for Robotic Planning

World models letting robots predict outcomes before acting are getting real traction. This one uses object-centric representations to predict multi-step physics, which could cut down the sim time you need.

Unpaired Manipulation: Learning From Unstructured Video Feeds

Another step toward getting manipulation policies from YouTube-scale video, not curated teleop datasets. If this works in the wild, it's a major unlock for teams that can't afford to collect their own pairs.

Constrained Optimization for Whole-Body Humanoid Control

Real-time whole-body control that respects joint limits and contact constraints. The kind of thing that actually ships on hardware.

VLM-Driven Dense Rewards for Long-Horizon Manipulation

Using vision-language models to create dense reward functions without hand-tuning. Could be a levers for RL teams who hate reward engineering.

Federated Policy Learning Across Heterogeneous Robot Fleets

Federated learning applied to robot fleets lets you train across heterogeneous hardware without pooling data. Particularly useful if you're running multi-site deployments.

Online Domain Adaptation for Sim-to-Real Policy Transfer

Tackling the sim-to-real gap directly with online adaptation during deployment. Cuts down on the painful fine-tuning loop.

Hierarchical LLM Planning with Composable Motor Primitives

Hierarchical task decomposition with LLMs as the high-level planner, learned skills as the low level. The bet is you can chain modular primitives with natural language.