Robot Adapters, Universal IK, and Video Foresight

Robot Foresight from Video Demos

This paper extracts actionable knowledge graphs from video demos to let robots foresee the consequences of their actions. It is a solid step toward robots that learn the cause-and-effect of manipulation tasks from just watching, without massive labeled datasets.

Modular Adapters for Robot Foundation Models

Instead of fine-tuning a whole foundation model for every new robot, these adapters let you plug in hardware-specific behavior on top of a general policy. This tech moves the goalpost on getting a single model to run on every piece of hardware you own.

Scalable Sim-to-Real for Manipulation

Closing the sim-to-real gap is still the biggest pain point in robotics, and this framework promises to do it at scale without real-world labels. If it holds up, your simulation spend gets a serious multiplicative boost in value because it actually transfers.

Universal IK from Transformers

By treating inverse kinematics as a sequence generation problem, this model handles highly kinematic diverse platforms with a single architecture. This is practical leverage for anyone building controllers for modular or reconfigurable robot arms.

Cooperative Skill Learning for Quadrupeds

Getting quadrupeds stacked and coordinated is a nightmare, so this contribution to tightly coupled multi-agent control is useful. Real-world outdoor experiments suggest this is closer to deployable wild robotics than just another benchmark.