Data Efficiency Meets Generalization, World Models Get Practical, and Robot Perception Tightens

Research

World Models for Robot Planning Move Beyond Toy Domains

World models have been hyped for years but mostly stayed in simulation. This work pushes them closer to real-robot decision-making, which is the leverage point that actually matters for deployment.

Learning Manipulation Policies With Less Data via Better Representations

Data hunger is the biggest bottleneck in robot learning right now. This paper tackles it head-on with representation learning that squeezes more signal from fewer demonstrations, which is exactly what production teams need.

Sim-to-Real Transfer Gets More Reliable for Contact-Rich Tasks

Contact-rich manipulation remains the hardest sim-to-real gap. The methods here tighten that transfer, which means fewer real-world hours needed to validate a policy before shipping.

Perception-Action Coupling Tightens for In-Hand Manipulation

In-hand manipulation demands tight perception-action loops and most systems still have too much latency. This work narrows that gap, which matters for any task requiring continuous dexterous control.

Language-Guided Robot Control Gets More Grounded in Physics

Language-conditioned policies often hallucinate actions that violate physics constraints. Grounding them in actual dynamics models is a practical step toward commands that work on the first try.

Benchmark

New Benchmark Exposes Gaps in Robot Generalization

Most robot benchmarks reward narrow overfitting. This one tests actual out-of-distribution generalization, which is the metric that predicts whether your demo video translates to a real warehouse.