Simulation Speeds Up, Affordances Learn to Ground, and Robots Get a Grip on Reality

Faster Physics: Accelerated Simulation for Robot Learning

This paper cuts the compute bill for training by speeding up physics simulation. Less time waiting for sim means more time iterating on real-world policies.

Grounded Affordances: Learning What Robots Can Actually Touch

Affordance prediction is finally getting grounded in physical reality, not just pixels. This means your robot can better predict what it can actually grab or push before it tries.

Tactile Sensemaking: Robots Get a Grip on Contact

New methods for interpreting tactile data are letting robots understand contact dynamics during manipulation. This is the difference between fumbling with an object and handling it with confidence.

Sim-to-Real Transfer Gets a Boost from Better Domain Randomization

Closing the sim-to-real gap is all about smart randomization. This work shows how to tweak simulation parameters so policies transfer more reliably to the messy physical world.

Structured Policies for Complex Manipulation Tasks

Instead of end-to-end black boxes, this research breaks down complex manipulation into structured, learnable sub-tasks. This makes debugging easier and improves success rates on hard jobs.