Introducing Waddle Labs: Claude Code for robots. Connect our API to your robot and enter a prompt, then our agents write code to achieve the task in 20 minutes. Yiding Song (Vincent) Waddle Labs

Robot models like VLAs are difficult to steer, rely on huge quantities of data, and don't yet generalize across embodiments and settings. On the other hand, LLMs are steered by conversation, transfer without finetuning, and extremely powerful at reasoning.

Our hypothesis is that these capabilities can be transferred into robotics by using LLM agents to control robots. Our agents decompose goals into subtasks, then complete each one by viewing camera feeds and writing control code. The output is a program that you can iterate on.

Using agents unlocks three capabilities: 1. Agents are generalists. They work with any robot and environment without new data. 2. Agents excel at long-horizon planning, and can re-plan on failure. 3. An agent can spawn subagents to coordinate an entire fleet of robots.

Waddle's agents enable new forms scaling. Our agent first wrote the “fold_grasp” skill while flipping a package; another agent later used it to fold a shirt. Over time, agents accumulate a library of skills, enabling future agents to compose them and achieve harder tasks.

As foundation models improve, so do our robots. We evaluated Opus 4.8, Fable 5, and GPT 5.6 Sol on a suite of tasks: larger models with larger thinking budgets produce better policies. More standardized benchmarks are needed to rigorously evaluate LLMs at robot control.

Read our technical blog: To request early access to the agent API, please reach us via [email protected]

@yiding_song @theWaddleLabs don’t tell me this simply loads an existing robotics model and runs it to fold the shirt and if it doesn’t, then i wonder if this type of code generation system could ever fold less cleanly laid out shirts quickly

@yiding_song @theWaddleLabs All the demos here were code-as-policy. No robotics models. Handling extreme-diversity tasks like crumpled shirts is tricky with code. But we're working on this and see encouraging signs (eg. programmatically grab corners of shirt, then flatten by whipping it in the air)

@yiding_song @theWaddleLabs The tricky part is how the agent will know when the shirt is flattened or not.

so agents calling more traditional robotics and perception stack tools? i’m fairly certain unless it pulls a learned model, it will not get good throughput or success rate for a number of tasks however, I do think there’s a subset of tasks where success rates/throughput aren’t needed. Here code as policies approach might work, exciting to see how this develops

@yiding_song @theWaddleLabs One use for code-as-policy has been collecting data that we then train models with. Our teleoperators were worse than code in terms of data quality.

@yiding_song @theWaddleLabs Right now is this just a wrapper? Because Claude Code is already pretty good at ssh’ing into hardware and learning the control scheme, then doing random activities until it converges on stuff the LLM understands as useful (and this capability naturally scales with newer models)

@yiding_song @theWaddleLabs Back in Feb we tried using raw CC on a robot. It was... not fun because Claude overthinks. It calculated quaternions for simple motion. We try to step out of the way of agents as much as possible, while steering them with how we design the tools they can use.

@yiding_song @theWaddleLabs Makes sense. A zero-shot can do cool stuff but only get so far, and will maybe eventually cover the surface area but expensively. Looking fwd to seeing how you can orchestrate and eval this toward something useful!

@yiding_song @theWaddleLabs We wrote about our approach here: Curious what you think? You're right eval is a big pain

@yiding_song @theWaddleLabs coolest thing about this is you’ve proven that it doesn’t take 10k hours of video to train a robot how to fold laundry. all those junk data startups will hopefully be leapfrogged :)

@yiding_song @theWaddleLabs Models will eventually get there with enough data, but we shouldn't rely on this data-scaling approach only (now or in the future).

@yiding_song @theWaddleLabs jarvis, fold my clothes for me

@yiding_song @theWaddleLabs This is very cool!

@yiding_song @theWaddleLabs congrats guys this is sick!!

@yiding_song @theWaddleLabs Brilliant! I recently connected AT Agents to ROS2 via @Agenticros. Let me know if it adds any value...

@yiding_song @theWaddleLabs @AgenticROS Wow there’s a lot of infra u’ve built - lemme read this more

@yiding_song @theWaddleLabs So cool. What’s the most complex task you guys got working with this setup? 👀

@s_haaris25714 @yiding_song @theWaddleLabs In terms of complex : effort ratio it was probably the flip package task. It took less time than expected and worked wonderfully

@yiding_song @theWaddleLabs come to yc paper club on Wed!

@yiding_song @theWaddleLabs Congrats!!

@yiding_song @theWaddleLabs this is unbelievably sick guys

@yiding_song @theWaddleLabs Thanks Samantha (and congrats!!) GPT 5.6 and Fable are very close in our evals… Excited to see how they progress

@yiding_song @theWaddleLabs but do @alex_brussell and i make you feel great or what!!?!?!?!

@DozenDucc @yiding_song @theWaddleLabs I’ve heard rumors that we might be releasing an even smarter, more capable model soon 👀

@strimblez @yiding_song @theWaddleLabs 👀More evals to run?

@yiding_song @theWaddleLabs @sincethestudy Have you guys tried anything like this with []bots?

@yiding_song @theWaddleLabs this is amazing!

@yiding_song @theWaddleLabs It's pretty much NVIDIA'S playbook but hardware agnostic and much faster/better. Awesome !! Which company would you like to be acquired by ? Microsoft ?? Google ?!

@yiding_song @theWaddleLabs Yea I agree we'd love to acquire Microsoft and Google

@yiding_song @theWaddleLabs We'll see :)

@yiding_song @theWaddleLabs so voyager on robotics with a nice set of starter tools i love it

@yiding_song @theWaddleLabs I know nothing about robotics, always felt specialized models were wrong Super exciting direction

@yiding_song @theWaddleLabs coolest demo ever, congrats!!

@yiding_song @theWaddleLabs v cool. great work guys!

@yiding_song @theWaddleLabs this is a really cool way to bootstrap, but does it permanently need an LLM in the loop or can i just iterate on the program it generates using existing RL techniques to go the last mile??

@yiding_song @theWaddleLabs You can use RL! In fact we've been trying out residual RL on top of a LLM-generated program for precise tasks like plugging in USB

@yiding_song @theWaddleLabs really cool! we are now getting robotics harnesses.

@yiding_song @theWaddleLabs I like that the agents build a reusable skill library over time. Does the system still work reliably when you switch to a completely different robot embodiment without any extra fine-tuning?

@yiding_song @theWaddleLabs Some robots have... unique physical characteristics. Maybe they sag under load, maybe they are less precise, or their native IK solver is inadequate. It could take time to handle these exceptions, but after that it's all handed to the agents

@yiding_song @theWaddleLabs The 20-minute prompt-to-code loop is impressive. The harder question for real deployments: when an agent writes control code that acts on hardware, what catches a bad action before it runs? Rollback is cheap in software, expensive when a robot has already moved.

@yiding_song @theWaddleLabs Good question - a lotta our work has gone into this. One thing we are working on: can we use sim / world models to evaluate programs before they run irl?

@yiding_song @theWaddleLabs That's the right direction. The open question for production is coverage — sim catches the failure modes you modeled, but the expensive ones are usually the edge cases nobody simulated. Do you pair it with a hard runtime boundary as a fallback, or lean fully on the sim?

@yiding_song @theWaddleLabs Super cool

@yiding_song @theWaddleLabs Congrats on the launch!

@yiding_song @theWaddleLabs Insane results!

@yiding_song @theWaddleLabs so amazing!

@yiding_song @theWaddleLabs Wow this is awesome! Congratulations team.

@yiding_song @theWaddleLabs Looks nice

@yiding_song @theWaddleLabs holy shit congrats hanming & yiding!!!!

@yiding_song @theWaddleLabs Thanks Sandra!

@yiding_song @theWaddleLabs this is insanely sick. done a bunch of smaller-scale llm agent + ftc/frc builds on my own, would kill to get my hands on hardware like this with you guys. let me know if you ever need interns

@yiding_song @theWaddleLabs THIS is cool.

@yiding_song @theWaddleLabs Congrats guys!

@yiding_song @theWaddleLabs congrats on yc in advance bro this is amazing

@yiding_song @theWaddleLabs A robot cannot be rolled back. Bad code gets reverted, a bad motion does not. Curious what sits between the generated code and the actuator, because that gate is the whole product.