Hanming Ye

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 @theWaddleLabs

Every AI company should own its data, its models, and the research lab that keeps making them better. Today, @thomasboser and I are launching @hiloopai to make that possible. The frontier advantage isn’t access to a model. It’s the research organization continuously improving it. Your product already generates the raw material for better intelligence: proprietary data, production feedback, evaluations, and domain expertise. Very few teams have the research capacity to turn those assets into better models. Hiloop builds and operates that capability with you. Bring us the model your product depends on, the data that makes it different, and an evaluation that defines success. We reproduce your baseline and run an autonomous research campaign against it. Agents pursue competing hypotheses in parallel, build on previous results, and promote only improvements that survive verification. Our researchers validate the winners and bring them into production. The work can run hosted or inside your environment. You retain control of your data, and the resulting models, evaluations, and research artifacts are yours. We’re starting with model training, post-training, and inference optimization, where progress is measurable and the value is immediate. But this isn’t a one-off model improvement. It’s a persistent research capability that begins each campaign with everything learned from the last. Over time, the lab accumulates research memory and improves its own tools, evaluations, and agents. The process used to create better intelligence gets better itself. That’s infrastructure for recursive self-improvement. If your product depends on a model and an important metric has stopped moving, bring us the model you can’t make better. We're much better at research than making videos! Reach out to us: founders@hiloop.ai

Introducing Rook by @Hop_Aero (YC S26): an autonomous hypersonic cargo rocket that delivers 550 lbs up to 450 miles in ~15 minutes. It launches from a standard 40-foot shipping container and lands on unprepared surfaces—no runways or fixed infrastructure required.

Putting LLM brain on robots -> 4x SOTA with no extra training. I’ve been very surprised by how well this works. The time for agents running on robots is coming soon

@NoahHWeiss I saw your work on LinkedIn back then! Was it called "farm" or smth like that?

This is really cool. These guys did something special with their harness! I tried something similar in May, and the LLMs then couldn't figure out how to coordinate basic picking and placing actions. This new sol/Fable gen has massively improved in this area. I did some testing within codex for 5.6sol xHigh, which spent about 4 hours building its own harness and roughly 1.5 hours executing for the clip below. I also tested Fable xHigh in claude code, which took about 2 hours to set up but eventually got stuck in a loop after 3 hours during its attempt. Both seemed to improve dramatically over the course of the run; I believe the chat log drove RL-like self-improvement. Raw model capability is finally crossing the threshold where spatial reasoning works, but it's still some way from being competitive with VLAs on inference speed + cost. GPT 5.6sol xhigh "👇

@myanvoos Curious what challenges you ran into? We prob faced the same things initially

This is really really cool and reminds me of a personal project back in June when I had Fable pilot my SO-101 arm to pick up cubes, with explicit instructions for it to write policies-as-code to do the task. There were some challenges with this ofc which makes me curious how these guys are approaching it

@Selen7005717917 @yiding_song @theWaddleLabs waddlelabs.ai/research gives an overview. At the bottom it also links to related papers!

@DozenDucc @yiding_song @theWaddleLabs Are there any blogs or papers about this work?

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 @theWaddleLabs

Sometimes code-as-policy works best (eg. repetitive tasks in factories), sometimes robot models work best (eg. walking a humanoid, controlling 5-fingered hand). People over-index on models, but we still need em. What's surprising is that code-as-policy and LLM controlling robots is far more capable than expected. Ths should unlock faster deployments and automated data collection.

@DozenDucc @yiding_song @theWaddleLabs Great work, would using code-as-policy wrapped over sophisticated VLAs be better than using frontier LLMs?

@artofelection @yiding_song @theWaddleLabs What task do you think we should work on

@DozenDucc @yiding_song @theWaddleLabs They need to stop with the folding t shirt. Thats the least of a problem no one cares about except lazy nerds

@hatebunnyplzzz @yiding_song @theWaddleLabs Cap-X is a foundational paper for code-as-policy! We didn't know "code-as-policy" was a thing in research when we started. But once we read about Cap-X it helped us design some of our primitives (see waddlelabs.ai/research)

@DozenDucc @yiding_song @theWaddleLabs is the idea inspired by Cap-X study from NVIDIA? looks impressive Interested to see what is the performance ceiling on some close loop evaluations including approaches involving world models as well. or even simulation like RoboCasa365

@DJiafei Also using TopReward and Robometer to judge rollouts :)

I see MolmoAct2 YAM setup.

@DJiafei Yes! Models beget users beget data begets models

@tomcocobrico @NickADobos Nah thanks Jeffrey you raise a very valid point (also its X). We've been using code-as-policy to collect tons of data internally - let's see where this goes

@DozenDucc @NickADobos Sorry didn’t wanna come across as an asshole here. I am sure it’s really exciting work. I wonder if this could be used scale the generation of training data instead of direct application (as it does sound quite expensive to run) Also robotics is a blind spot for me 😬

Wait if legit this is a huge deal. Did they solve robotics?!?!

It sounds inelegant to burn a million tokens to fold a shirt. But isn't it cool that you can do this at all with a big next-token predictor?

@NickADobos solve? you mean how to burn millions of tokens to fold a shirt? sure

@NickADobos What's surprising is that code-as-policy and LLM controlling robots is far more capable than expected. I expect this to unlock faster deployments and automated data collection.

@NickADobos I don't think so. There’s no one path to robot intelligence. Sometimes code-as-policy works best (eg. repetitive tasks in factories), sometimes robot models work best (eg. walking a humanoid, controlling 5-fingered hand). People over-index on models, but we still need em.

@Sujeetsoni123 @yiding_song @theWaddleLabs There's hacks to get these models to work better but defo much room for improvement

@DozenDucc @yiding_song @theWaddleLabs Got it, in that perspective, yes it's better than turn based model but can't we hack around by forcing the frequency to much higher or using multiple async agent? Also do you think current llm understand the physics better or there's still gap or you didn't observe yet ?

@Sujeetsoni123 @yiding_song @theWaddleLabs LLMs definitely suck as physics. Maybe the root cause is that they suck at perception. We tried many models (every gemini, Nvidia Cosmos, you name it) and they all hallucinate an unbearable amount on videos.

@BrutalCaeser @yiding_song @theWaddleLabs In all the demos in the video there were no robot models (eg. ACT, VLAs). This goes to show that code-as-policy works in diverse situations. I believe that robot models will be crucial for dexterous tasks, it's just that they're not the only way towards physical intelligence.

@DozenDucc @yiding_song @theWaddleLabs Love the legibility angle @DozenDucc . But code-as-policy is only as good as its primitives, and the hard part (contact-rich dexterity) lives inside them. You still train ACT and call VLAs as tools. Isn't that moving the robot-data problem down a layer, not removing it?

@BrainsAndTennis The voyager paper was a substantial inspiration for our skill library design!

voyager on robots via a generalist frontier model > VLAs arxiv.org/abs/2305.16291 i love how bitter this is

@Sujeetsoni123 @yiding_song @theWaddleLabs I thought about duplex models as a way to improve accuracy actually. Because turn-based models miss changes in the environment unless they call perception tools intentionally. So full-duplex would make it easier to catch things like wind blowing over a water bottle.

@DozenDucc @yiding_song @theWaddleLabs Interesting, yes duplex vision model can reduce latency a lot. I'd have focus on accuracy first instead of latency because there are several ways to reduce latency later but accuracy will be the moat. Combination or world model and LLM can give more physics aware output I think.