Waddle Labs

Waddle Labs offers an API that lets developers connect a robot, type a prompt, and have an agent produce a working policy for the task. The system is built for robotics teams who currently rely on teleoperation data and vision-language-action models to teach new behaviors. Rather than train a large end-to-end model to map pixels and instructions to motor commands, the agent looks through the robot's cameras and then writes code, uses skills, or calls specialist models such as VLAs, producing a program that can be edited afterward. The launch matters because it reframes robot learning as a code generation problem that scales with frontier LLMs. Planning is handled by the reasoning model rather than the policy, so the agent decomposes long tasks into stages, verifies intermediate outcomes, and re-plans on failure, spawning subagents to coordinate multiple robots when needed. Agents also build a library of parametrized skills like grasping, aligning, and folding, and reuse them across tasks (a fold_grasp skill first created to flip a package was later adapted to fold a t-shirt and a towel). The team pitches four practical uses: direct deployment of the generated program on hardware, supervised data collection, fast baseline policies for new tasks or robots, and autoresearch where an agent runs its own hardware trials against a hypothesis and a budget. Waddle Labs is built by Hanming Ye and Yiding "Vincent" Song, both Harvard students who, despite wanting to pursue Waddle full-time, are staying at Harvard because their international student visas require it . Their positioning is that policy quality tracks model quality: across a manipulation task suite evaluated on Opus 4.8, Fable 5, and GPT 5.6 Sol, larger models given larger budgets produced better policies, with only Fable 5 and GPT 5.6 at xhigh thinking accomplishing harder tasks like folding a t-shirt . Early access is gated through a request form on the company's site.