Simulation Gets Cheaper, Affordances Get Richer, and Manipulation Learns to Wait

Cheap Sim, Real Robots: Scaling Robot Learning Without Breaking the Bank

This paper tackles the cost and compute barrier to scaling sim-to-real. If you're bootstrapping a robot stack, this is the kind of leverage that lets you iterate on policy without renting a GPU farm.

Affordances That Actually Think About Space

Goes beyond simple grasp-point detection to reason about spatial relationships and object context. This is a step toward robots that understand the 'why' behind a grasp, not just the 'where'.

The Strategic Pause: Learning When Not to Act in Manipulation

Introduces a mechanism for a robot to intentionally wait and gather information before acting. It's a counterintuitive but powerful idea—patience as an active policy choice.

Generalizing Manipulation Skills Across Novel Objects

Focuses on transferring learned manipulation policies to objects the robot has never seen before. This kind of generalization is the difference between a demo and a product.

Putting Physical Constraints Back Into Robot Learning

Re-introduces explicit physics priors and constraints into the learning loop. It's a back-to-basics move that can prevent brittle, overfit policies and improve sample efficiency.