Waddle Labs

Waddle Labs is a two-person American startup, founded in 2025 and based in Boston, that sells an API through which large language model agents write, run and revise the control code that drives a customer's robot. It launched on 27 July 2026 with a landing page, a 91-second video and one technical blog post, and belongs to Y Combinator's Summer 2026 batch.^[33] Founders Yiding Song and Hanming Ye pitch it as "Claude Code for robots": connect a robot, type a prompt, and the agents return a working program in about 20 minutes. Every capability claim below is the company's own; as of 31 July 2026 it has published no paper, no success rates, no code and no independent evaluation.

Scroll sideways for more →

Company and founders

Y Combinator's directory lists Waddle Labs as an active Summer 2026 company, founded in 2025, based in Boston, with two employees, no open roles, group partner Ankit Gupta, and the one-line description "AI agents that autonomously control robots by decomposing tasks into subtasks and generating executable code."^[1] YC publishes its standard deal terms (125,000 US dollars on a post-money SAFE for 7 percent, plus 375,000 dollars on an uncapped SAFE with a most-favored-nation provision), but nothing about Waddle Labs' own financing has been disclosed.^[2]

Both founders are undergraduates. Song's personal site describes him as a Harvard student in computer science (mind, brain and behavior) and physics who researches how Transformers perform program induction through in-context meta-learning, affiliated with the Gershman Lab and the Kempner Institute.^[3] Ye's Google Scholar profile lists Harvard University and algebraic topology, and records a 2023 paper on the stable Picard groups of the exterior algebras E(n) credited to MIT, a 2025 NeurIPS workshop paper on steering diffusion policies with value-guided denoising, and a joint 2026 preprint with Song on how model capacity determines grokking.^[4][5] His X profile reads "math @ harvard | rsi '23".^[7] The two met at the Research Science Institute, a six-week program for rising high-school seniors hosted at MIT and administered by the Center for Excellence in Education; its mathematics students are matched with graduate mentors by the MIT mathematics department, which is why Ye's topology paper carries an MIT line rather than a current MIT affiliation.^[6] RuntimeWire, a trade outlet that drafts copy with a language model under editorial review, reports both founders are in Harvard College's Class of 2028.^[8]

What the company says the product does

The launch post, "Introducing Waddle: agents that control robots," argues that the dominant recipe in robot learning, training an end-to-end model that maps camera frames plus an instruction to motor commands, needs too much robot-specific data, produces models that are hard to steer, and does not generalize across embodiments. Waddle's alternative puts an agent above the control layer: it decomposes a goal into subtasks and completes each by viewing camera feeds, writing control code, and calling models such as vision-language-action models as tools. The output is a program the operator can run, read, and revise by talking to the agent rather than by collecting a new dataset.^[9]

Generated programs sit on a three-level hierarchy:

Scroll sideways for more →

The company claims three capabilities, all said to be demonstrated on real hardware: the agents are generalist and work with any arms, grippers and camera setups without new data collection; planning is done by the reasoning model rather than the policy, so the agent stages long tasks, verifies intermediate outcomes and re-plans on failure; and controlling more robots requires only connecting more agents, with a master agent spawning subagents. It says its skill library transfers across tasks, giving one example: an agent created fold_grasp while trying to flip a package, and a different agent later adapted it to fold a towel and a t-shirt.^[9]

Three use cases are shown: a policy for "place one microswitch inside each slot" produced in 20 minutes; an agent that ran roughly a thousand randomized Lego pick-and-place repetitions overnight and trained an action chunking transformer (ACT) policy from scratch on the resulting data; and an agent resetting a scene between trials while a human tuned a policy overnight. The company also reports an internal comparison of Claude Opus 4.8, Fable 5 and GPT-5.6 Sol on its own manipulation suite: larger models with larger reasoning budgets produced better policies, all three handled "pick up the lego," and only Fable 5 and GPT-5.6 at the highest thinking setting folded a t-shirt.^[9]

What has not been shown

There is no preprint or peer-reviewed publication from Waddle Labs. The only success-rate result in the launch post is a curve against cost per task, with no tabulated numbers, task list, trial counts, or error bars. The 20-minute figure attaches to a single task (microswitch insertion) with no stated task distribution, success threshold, or number of prior attempts. The API is distributed by request rather than as a self-service product, so no outside party can reproduce any of it, and no third-party evaluation exists. Waddle Labs itself notes that results in this area are hard to compare because tasks, robots, and success criteria differ across labs, and calls for a shared benchmark for agent-controlled robots.^[9] Its GitHub organization, created in June 2025, holds three public repositories, all forks of existing open-source projects: the dora dataflow middleware, a Joy-Con teleoperation package, and an XLeRobot repository derived from LeRobot. No Waddle Labs code of its own is public,^[10] and the robots in the launch video are not identified by model anywhere on the company's site.

LLM-generated code as a control paradigm

The category Waddle Labs is entering is usually called code-as-policy: instead of a network emitting joint torques, a language model writes a program that calls perception and control routines, and the program is what runs on the robot.

Code as Policies (Liang et al., Google Robotics, ICRA 2023) established the pattern: given few-shot examples of command-and-program pairs, code-writing LLMs re-composed API calls into new policy code that showed spatial-geometric reasoning, generalized to new instructions, and mapped vague terms such as "faster" onto concrete velocities. Its central mechanism was hierarchical code generation, recursively defining undefined functions.^[11] ProgPrompt (Singh et al., NVIDIA, ICRA 2023) worked the planning side, prompting the model with program-like specifications of available actions and objects so that plans stay executable instead of drifting into free-form text the robot cannot run; it reported state-of-the-art success on VirtualHome household tasks and ran on a physical arm.^[12] By Waddle Labs' own reading of this literature, the limitation of the early code-writing line was that the program was generated once per instruction, with revision, where it existed, coming from human corrections.^[9]

Two 2022 systems supplied grounding. SayCan (Ahn et al., Google, CoRL 2022) paired an LLM's semantic knowledge with value functions attached to pretrained skills, so the model could only propose actions feasible in the current scene.^[13] Inner Monologue (Huang et al., CoRL 2022) closed the loop with natural-language feedback (success detection, scene description, human corrections), which significantly improved instruction completion on tabletop rearrangement and long-horizon kitchen mobile manipulation, with no additional training.^[14] Text2Motion (Lin et al., Autonomous Robots 2023) added geometric feasibility planning across skill sequences and reported 82 percent success on long-horizon sequential manipulation against 13 percent for prior language-based planners, a measure of how brittle the early planners were.^[15] RoboCodeX (Mu et al., 2024) turned code generation into a tree-structured multimodal decomposition carrying affordance and safety constraints.^[16]

A parallel branch had the model write objectives rather than programs: Eureka (Ma et al., NVIDIA and the University of Pennsylvania, ICLR 2024) ran evolutionary optimization over reward function code with GPT-4, beating human-designed rewards on 83 percent of 29 reinforcement learning environments spanning 10 robot morphologies,^[17] and DrEureka (Ma et al., RSS 2024) extended it to sim-to-real transfer by having the model write domain-randomization distributions too.^[18] The skill-library idea behind "compositional generalization" comes from Voyager (Wang et al., 2023), an LLM agent in Minecraft with an automatic curriculum, an ever-growing library of executable skills and a prompting loop driven by environment feedback, execution errors and self-verification. It obtained 3.3 times more unique items and reached tech-tree milestones up to 15.3 times faster than prior systems, and its skills transferred to a fresh world.^[19] Waddle Labs cites Voyager as the pattern its library follows.^[9]

Three 2026 papers are the immediate context for the launch:

Scroll sideways for more →

Anthropic's "Claude Plays Robotics" (9 July 2026) is the clearest published measurement of where the paradigm stands. Testing Claude models on classic control problems, a simulated Unitree Go2 and G1, a 7-DoF Franka Panda and a physical Go2, it found models mostly fail when they must drive joints directly (0 to 5.5 percent task completion) but do markedly better supervising a pretrained controller. Interface design mattered more than model scale in one case: adding a cursor tool, a small red X on the gripper camera the model can move and query for the object and distance at that point, raised success on a 10-task subset from 6 percent to 32 percent for one model.^[23] That gap between direct and tool-mediated control is the commercial premise of a company like Waddle Labs.

How this differs from the VLA approach

The competing approach trains an end-to-end vision-language-action model: one network that consumes camera frames and an instruction and emits actions. RT-2 (2023) showed web-scale vision-language pretraining transfers into robotic control,^[24] pi-zero (Physical Intelligence, 2024) added a flow-matching action head trained across multiple robot types,^[25] and Gemini Robotics 2 (Google DeepMind, 30 July 2026) extended the family to whole-body humanoid control, 22-DoF multi-finger hands, and robot-to-robot collaboration, reporting multi-finger success rates of 32 to 92 percent by task and an on-device variant said to adapt to a new embodiment in hours from fewer than 200 examples.^[26] Figure's Helix, NVIDIA Isaac GR00T, and Skild AI's Skild Brain sit in the same family of robot foundation models.

The trade is fairly clean:

Scroll sideways for more →

The asymmetry has a simple cause. A program is a discrete object: it captures what is discrete about a task (which object, which order, what to check next) and cannot capture what is continuous and reactive (how hard to press, how to recover from a slipping grasp mid-motion) unless a primitive already implements it. That is why the strongest results in this line come from tasks with clear symbolic structure, and why Waddle Labs' own architecture keeps calling VLAs as tools rather than replacing them.

The "Claude Code for X" framing

A coding agent is a loop, not a model: an LLM given file-editing and shell tools, run repeatedly against a codebase, allowed to execute what it writes, read the errors, and try again. Claude Code is the reference implementation and the source of the analogy. Robotics is a plausible target because the models are already competent program synthesizers and the robot supplies the ingredient a coding agent needs, an executable environment that returns feedback. The disanalogy is what a failure costs. A failed compile is free; a failed grasp can break a fixture, and the agent's next turn arrives after the damage rather than before it.

Open questions

Executing generated code on physical hardware. Nothing published describes what sandboxing, force limits, workspace bounds or human oversight sits between an agent-written program and a moving arm. The fixed platform primitives are a natural control point, since they are the only things skills can call, but Waddle Labs has not documented what they refuse to do. This is the central robot safety question for the category and it is unanswered.

Validation. Running on real hardware rather than in simulation sidesteps the sim-to-real gap but substitutes a sample-size problem: real-robot trials are slow, and no trial counts are given for anything.

Scaffolding. The CaP-X finding, that code-as-policy agents lean heavily on human-crafted abstractions and degrade when those are removed, is exactly what a "works with any robot without new data" claim has to answer, since the platform primitives are such an abstraction.^[20]

Competitive and adjacent landscape

Scroll sideways for more →

The distinction that matters is where the intelligence sits. Physical Intelligence, Skild, Covariant and the Gemini Robotics family put it inside a trained policy. Wandelbots puts it in a portable abstraction layer that humans program. Waddle Labs puts it in an agent that writes against such a layer, and says it plans to train its own models on the six months of robot interaction and intervention traces it has accumulated.^[9] Its nearest neighbours are therefore not the foundation-model labs but the research systems: ASPIRE at NVIDIA and VIA at Stanford do the same thing without a company attached, and CaP-X exists to measure it.^[20][21][22] The idea has a real lineage, and 2026 has produced independent evidence that it works in a laboratory; whether it holds up on a customer's hardware is untested.

See also

• Vision-language-action model
• Robot foundation model
• Coding agent
• Embodied AI
• Robot learning
• Physical AI
• AI agent
• Y Combinator

References

1. ^"Waddle Labs: Agents for robotics control." Y Combinator company directory. Accessed 31 July 2026. ycombinator.com/...waddle-labs
2. ^"The YC Deal." Y Combinator. Accessed 31 July 2026. ycombinator.com/deal
3. ^Yiding Song (Vincent), personal site. Accessed 31 July 2026. yiding.rocks
4. ^Hanming Ye, Google Scholar profile. Accessed 31 July 2026. scholar.google.com/citations
5. ^Yiding Song and Hanming Ye. "Model Capacity Determines Grokking through Competing Memorisation and Generalisation Speeds." arXiv:2605.09724, 10 May 2026. arxiv.org/...2605.09724
6. ^"RSI: Research Science Institute." MIT Department of Mathematics. Accessed 31 July 2026. math.mit.edu/...rsi
7. ^Hanming Ye (@DozenDucc). "Introducing Waddle Labs: Claude Code for robots." X, 27 July 2026. x.com/...2081802609128788393
8. ^Ryan Merket. "Waddle Labs launches AI agents that write control code for robots." RuntimeWire, 27 July 2026. runtimewire.com/...agents-write-robot-control-code
9. ^Waddle Team. "Introducing Waddle: agents that control robots." Waddle Labs Blog, July 2026. waddlelabs.ai/...introducing-waddle
10. ^Waddle Labs GitHub organization. Accessed 31 July 2026. github.com/waddlelabs
11. ^Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman et al. "Code as Policies: Language Model Programs for Embodied Control." arXiv:2209.07753, 16 September 2022 (ICRA 2023). arxiv.org/...2209.07753
12. ^Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu et al. "ProgPrompt: Generating Situated Robot Task Plans using Large Language Models." arXiv:2209.11302, 22 September 2022 (ICRA 2023). arxiv.org/...2209.11302
13. ^Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar et al. "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances." arXiv:2204.01691, 4 April 2022 (CoRL 2022). arxiv.org/...2204.01691
14. ^Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang et al. "Inner Monologue: Embodied Reasoning through Planning with Language Models." arXiv:2207.05608, 12 July 2022 (CoRL 2022). arxiv.org/...2207.05608
15. ^Kevin Lin, Christopher Agia, Toki Migimatsu, Marco Pavone, Jeannette Bohg. "Text2Motion: From Natural Language Instructions to Feasible Plans." arXiv:2303.12153, 21 March 2023 (Autonomous Robots, special issue on LLMs in robotics). arxiv.org/...2303.12153
16. ^Yao Mu, Junting Chen, Qinglong Zhang, Shoufa Chen, Qiaojun Yu et al. "RoboCodeX: Multimodal Code Generation for Robotic Behavior Synthesis." arXiv:2402.16117, 25 February 2024. arxiv.org/...2402.16117
17. ^Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani et al. "Eureka: Human-Level Reward Design via Coding Large Language Models." arXiv:2310.12931, 19 October 2023 (ICLR 2024). arxiv.org/...2310.12931
18. ^Yecheng Jason Ma, William Liang, Hung-Ju Wang, Sam Wang, Yuke Zhu et al. "DrEureka: Language Model Guided Sim-To-Real Transfer." arXiv:2406.01967, 4 June 2024 (RSS 2024). arxiv.org/...2406.01967
19. ^Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao et al. "Voyager: An Open-Ended Embodied Agent with Large Language Models." arXiv:2305.16291, 25 May 2023. arxiv.org/...2305.16291
20. ^Letian Fu, Justin Yu, Karim El-Refai, Ethan Kou, Haoru Xue et al. "CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation." arXiv:2603.22435, 23 March 2026. arxiv.org/...2603.22435
21. ^Runyu Lu, Yubo Wu, Ethan Kou, Letian Fu, Wenli Xiao et al. "ASPIRE: Agentic Skills Discovery for Robotics." arXiv:2607.00272, 30 June 2026. arxiv.org/...2607.00272
22. ^Hengyuan Hu, Priya Sundaresan, Jensen Gao, Dorsa Sadigh. "VIA: Visual Interface Agent for Robot Control." arXiv:2607.11119, 13 July 2026. arxiv.org/...2607.11119
23. ^"Claude Plays Robotics." Anthropic, 9 July 2026. anthropic.com/...claude-plays-robotics
24. ^Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen et al. "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." arXiv:2307.15818, 28 July 2023. arxiv.org/...2307.15818
25. ^Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi et al. "pi-0: A Vision-Language-Action Flow Model for General Robot Control." arXiv:2410.24164, 31 October 2024 (RSS 2025). arxiv.org/...2410.24164
26. ^"Gemini Robotics 2 brings whole body intelligence to robots." Google DeepMind, 30 July 2026. deepmind.google/...ole-body-intelligence-to-robots
27. ^"Physical Intelligence raises $400M for foundation models for robotics." The Robot Report, 4 November 2024. therobotreport.com/...undation-models-for-robotics
28. ^"Robotic software maker Skild AI hits $14B valuation." TechCrunch, 14 January 2026. techcrunch.com/...aker-skild-ai-hits-14b-valuation
29. ^"Amazon hires from AI robotics startup Covariant, licenses technology." About Amazon. Accessed 31 July 2026. aboutamazon.com/...amazon-covariant-ai-robots
30. ^"Unpacking Amazon's unique Covariant AI deal." The Robot Report, 9 September 2024. therobotreport.com/...que-covariant-ai-acquisition
31. ^"Dexterity picks up $95M in funding for container loading robots." The Robot Report, 12 March 2025. therobotreport.com/...g-container-unloading-robots
32. ^"Wandelbots NOVA: The Robot-Agnostic Operating System." Wandelbots. Accessed 31 July 2026. wandelbots.com/wandelbots-nova
33. ^"Introducing Waddle Labs: Agents that control robots." Waddle Labs, YouTube, 27 July 2026. youtube.com/watch
34. ^Open X-Embodiment Collaboration. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." arXiv:2310.08864, 13 October 2023. arxiv.org/...2310.08864