Policies Get Leaner, Navigation Gets Spatial, and Robots Learn to Generalize from Fewer Demos

Compact Policy Distillation for Real-Time Robot Control

This paper tackles the gap between large, high-performing policies and the lean models needed for real-time control on edge hardware. If you're building robots that need to think fast without a GPU farm, this is the kind of model compression work that makes it possible.

Semantic Spatial Reasoning for Long-Horizon Navigation

The key move here is grounding navigation not just in geometry, but in semantic understanding of spatial relationships. This is how you get a robot to reliably navigate a messy workshop or office, not just a clean hallway.

Learning Manipulation from a Single Teleop Demonstration

This pushes the few-shot boundary by extracting a reusable manipulation policy from just one demonstration. The real value is in drastically cutting the cost of teaching a robot a new task, which is currently a massive data bottleneck.

Tactile-Informed Policy Transfer Across Different Grippers

This research shows how to transfer tactile sensing knowledge between robots with different hand designs. It's a practical step toward creating a general 'sense of touch' that isn't locked to a specific piece of hardware.

Faster Simulation via Learned Physics for Robotic Assembly

They're using a learned physics model to speed up sim-to-real transfer for assembly tasks. For anyone building in simulation, this means you can iterate faster without waiting for slow, high-fidelity physics engines to render every millisecond.