| Project Weasel Project Weasel is a research initiative aimed at making Vision-Language-Action (VLA) models genuinely useful in the real world — focused on agents that take action. The main project, the W series, identifies and tackles the core problems VLAs face today, working toward agents with real autonomy. Sub-projects explore important VLA-adjacent topics. |  |
Main project |
 | W0: A dVLM-Based GUI Agent with Block-Diffusion Mode Steering Flagship of the W series: a dVLM-based GUI agent that keeps a frozen screen-language backbone and a block-discrete diffusion action head, training only a mode selector with critic-free RLVR (MSPO) for fast, long-horizon GUI control. |
Sub-projects |
 | Mean Collapse is a Serious Bottleneck of VLA We test the mean-collapse case on Ant-Maze. For a VLM-based VLA acting on the latest observation only, mean collapse turns out to be a major problem on POMDP inputs — and a small memory-based steering module on the frozen VLA solves it, raising the success rate above the MDP baseline. |
 | MineMate: Multi-Agent Emergent Capability with Humans Studying how cooperative behaviour emerges when humans are part of a multi-agent system. |