For the better part of six months, Yiding Song and Hanming Ye ran robots continuously. Around the clock, no breaks. This wasn't some academic exercise tucked into a lab at MIT or Harvard, where both had studied. This was production work—the kind that breaks things, surfaces edge cases, and generates the messy data that's supposed to teach machines how the physical world actually behaves.
The pair met years earlier at MIT's Research Science Institute, became roommates at Harvard, and then did what a certain breed of technically ambitious twentysomethings does: they started a company. In late July, Song and Ye walked out of Y Combinator's most recent batch with Waddle Labs, a platform they describe as "agents that control robots." The promise? Turn natural language into working robot code, as easily as GitHub Copilot autocompletes a function.
Well, almost as easily.
The pitch goes like this: Connect your robot to Waddle's API. Type what you want it to do—"place one microswitch in each slot," say—and the platform's agents will generate a runnable control policy in about 20 minutes. No sprawling datasets required upfront. No months spent training models in simulation. Just prompt, iterate, and deploy.
It's the kind of claim that sounds either transformative or wildly optimistic, depending on how many robotics demos you've seen fail in person.
Breaking Down the Task
What Waddle is actually doing sits somewhere between code generation and autonomous experimentation. The platform decomposes a task into subtasks, monitors camera feeds, writes control code, and spits out a program the robot can execute. Developers refine the policy by conversing with the agent—essentially debugging through dialogue.
Song and Ye have been stress-testing this loop on real hardware. Their July launch writeup describes agents autonomously collecting 1,000 LEGO pick-and-place cycles overnight, then training an ACT policy from scratch using that data. Early access is live on their site, though the company hasn't disclosed pricing. Or customers, for that matter.
The technical bet here diverges from the prevailing orthodoxy in robotics AI. Current state-of-the-art approaches—vision-language-action models, world action models—demand enormous datasets and often stumble when asked to generalize across different robot bodies. Waddle is inverting the problem. Instead of training a single foundation model to predict actions, they're using large language models to write and revise control code. Action models become tools the agent can call when appropriate, not the architecture itself.
The platform supplies primitives: perception modules, inverse kinematics solvers, trajectory planners. Agents compose these into parametrized "skills" like fold_grasp, which nest into larger programs and can be reused. It's a software engineering mindset applied to manipulation tasks, treating modularity as the unlock rather than scale.
Internal testing—conducted across models Waddle identifies as "Opus 4.8," "Fable 5," and "GPT 5.6 Sol"—showed that beefier models with larger "thinking budgets" handled complex manipulation like folding a t-shirt, while smaller models sufficed for simpler operations. Performance, in other words, scales with the foundation models underneath. Whether those models will keep improving at the pace Waddle needs is a different question.
The Founders
Song studied computer science and physics at Harvard, working at the Kempner Institute and the Gershman Lab before diving into the startup. He'd previously spent time at MIT's NSF AI Institute for Artificial Intelligence and Fundamental Interactions. Ye, who contributed mathematics research during his RSI days, authored a paper titled "The Stable Picard Groups of the Exterior Algebras E(n)" in August 2023 during his time at the MIT RSI—the kind of esoteric work that suggests someone comfortable with abstractions.
The company was founded in 2025 and entered Y Combinator's most recent batch. As of their public emergence, the team was still just the two founders, operating out of Boston. No outside funding beyond YC has been announced.
A Crowded Moment
Waddle's timing is either opportunistic or perfectly synced with a broader shift in how people think about robotics tooling. Weeks before the launch, Anthropic published research exploring how Claude could control robots across multiple abstraction levels—from raw torque commands to high-level steering instructions. The work demonstrated that foundation models could handle physical tasks but left a conspicuous gap: Where's the developer tooling?
That's the opening Waddle is sprinting toward. The team references projects like VIA (Visual Interface Agent for Robot Control), Code as Policies, VoxPoser, and Inner Monologue as adjacent efforts in a landscape that's evolving rapidly. Around the same time, Roboto AI announced agents for automating data workflows in what they call "Physical AI," while Agency Tool Company—also from the same YC batch—launched deployment tools claiming to be 20 times faster than traditional methods.
Perhaps more telling: Walden Robotics emerged from stealth in mid-July with $300 million in funding, a signal that capital is flooding into general-purpose robotics. The infrastructure layer for that wave, though? Still being built.
What Happens Now
Waddle's immediate roadmap focuses on refining interfaces for agents and establishing standardized benchmarks—the unglamorous work of turning research demos into production tools. Song and Ye plan to train more capable agent models using data harvested from their six-month robot marathon: intervention traces, failure logs, the accumulated detritus of systems colliding with reality.
Early access requests are open. No customers have been named publicly. The technical writeup includes demos and internal evaluations, but external case studies haven't appeared yet.
In a social media post aggregated in late July, someone described Waddle as "Claude Code for robots"—a framing that neatly captures the ambition. If the platform delivers, robotics developers won't burn weeks hand-tuning policies. They'll describe intent in plain language and let agents generate the rest.
Whether that vision survives contact with actual warehouses, factory floors, and all the chaos that entails is the test every robotics startup eventually faces. Song and Ye have spent half a year running robots nonstop. Now comes the harder part: convincing others to do the same.
