Waddle Labs

AI has advanced quickly, but robots still struggle outside carefully prepared settings. Making a robot useful often takes work specific to one machine, task, or lab. That makes progress difficult to reproduce and extend, and risks concentrating robot intelligence in a few large companies. We believe individuals should be able to adapt robots to their own needs. Our research aims to make that possible.

• Model Capability. Language models are not perfect at robotics control, but they can learn in context. We have seen Waddle's agents become markedly more reliable over time. How do we preserve what one agent learns and transfer it to another? How do we build a shared memory across agents? Our work draws on and extends research in continual learning.
• Inference Efficiency. We have shown that language models can solve robotics tasks in real time, on real robots, but latency remains a bottleneck. How can agents become faster with practice? How can we improve their efficiency while preserving the flexibility to adapt to new tasks and environments?
• Human-robot interaction. Robot intelligence must be built for people. We study how people can give robots tasks, see what they are doing, and correct mistakes. We design interfaces that make it easy and safe for individuals to collaborate with robots.