Research

GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning

Duke Physical AI and Robotics Lab

LLM-generated robot plans often look reasonable but can't actually be executed. GAVEL keeps an explicit graph world model of objects, relations, action preconditions and effects. Here it can roll forward LLM-proposed actions to catch and fix possible violations. The graph can also deterministically reason over likely object distributions and reorder future tasks as it observes more of the scene.

LLM Planning · World Models

NNiT: Width-Agnostic Neural Network Generation with Structurally Aligned Weight Spaces

Duke Cyber-Physical Systems Lab

A diffusion transformer that generates neural network weights directly, skipping training. Weights from a Graph HyperNetwork are structurally aligned, so they can be patch-tokenized like an image and generated at any layer width. NNiT generates working RL policies for architectures unseen in training (>85% success on ManiSkill3), and was tested on a Unitree Go2.

A generated policy doing the PushT task in ManiSkill.

Diffusion · Reinforcement Learning · ManiSkill · PyTorch