Waddle Labs

Waddle Labs develops an agent-based control layer for robots. Customers connect a robot to the company's API and issue a natural-language prompt; LLM agents then decompose the goal into subtasks and complete each one by viewing camera feeds, writing control code, and calling action models such as vision-language-action models (VLAs). The agent emits a program that the user can run and further iterate on through conversation with the agents, with the company stating a working policy can be produced in roughly 20 minutes.

The company frames its approach as an alternative to the two prevailing options in robot intelligence: large end-to-end VLAs or world action models, which it describes as data-hungry, hard to steer and not yet generalizing across embodiments and environments; and classical engineered approaches, which it describes as reliable and explicitly controllable but brittle and costly to rebuild for new setups. Waddle Labs positions LLM agents writing code as policy as a way to combine the generalization of the former with the speed and reliability of the latter, while retaining the ability to call action models as tools.

Stated use cases include rapid policy creation (for example, "write a program to place one microswitch inside each slot"), autonomous data generation for model training (the company reports having its agent pick and place LEGO bricks roughly a thousand times overnight and then autonomously train an ACT policy from scratch), and robotics auto-research tasks such as resetting a scene between overnight policy-tuning trials.

Founding story

Founders Yiding Song (Vincent) and Hanming Ye met at MIT's Research Science Institute, where Song worked on multimodal astrophysics models and Ye researched topology in the mathematics department. They subsequently became roommates at Harvard, working on machine learning research and robotics. Ye serves as Founder/CEO and describes an interest in how natural intelligence arises and in machines such as 3D printers and robots; Song describes an interest in systems that learn theories about the world with human-like efficiency. The company states a mission of enabling everyone to take part in building the robots it expects to exist in the future.

Business model

Waddle Labs delivers its agent system as a developer platform: robots connect through the company's API, agents run against the connected hardware, and the agents output programs the customer can execute and iterate on. A preview API is offered via sign-up on the company website or by contacting the founders directly.

Traction

The company reports internal use of its agents to create working policies in about 20 minutes, to generate training data (approximately 1,000 autonomous LEGO pick-and-place repetitions overnight followed by autonomous training of an ACT policy), and to run overnight auto-research routines. Team size is listed as two.

Latest developments

Waddle Labs is listed as an active Y Combinator Summer 2026 company and is offering preview access to its agent API.