Waddle Labs develops an LLM-agent framework that controls robots by writing and executing code instead of relying on end-to-end trained action models. Its agents decompose long-horizon tasks into subtasks, view camera feeds, and call action models like VLAs as tools, enabling generalist control across robot arms, grippers, and camera setups without data collection or fine-tuning. The system supports multi-agent coordination for concurrent robots and can generate training data, allowing users to create working policies in roughly 20 minutes.
Funding
Funding not disclosed
Founders
Product
Problem
Robot learning currently relies on training end-to-end models, such as vision-language-action models (VLAs) or world action models, that map camera frames and instructions directly to motor commands. This approach requires enormous amounts of robot data, produces models that are difficult to steer, and fails to generalize across robot embodiments and environments, making deployment to new hardware or tasks a slow, data-intensive process.
Solution
Waddle Labs provides an LLM-agent framework that controls robots through code-as-policy, where an agent decomposes a high-level instruction into subtasks and completes each by viewing camera feeds, writing executable control programs, and calling action models like VLAs as tools. Because the agent is an LLM agent, it inherits generalist capabilities, working with any robot arm, gripper, or camera setup without needing new data collection or fine-tuning. The agent handles long-horizon planning by reasoning through stages, verifying intermediate results, and re-planning on failure, while a master agent can spawn subagents to coordinate multiple robots concurrently. Users interact with the agent conversationally, iterating on generated programs until the robot performs the desired task.
Target Audience
Primary users are robotics researchers, automation engineers, and companies deploying industrial manipulators who need flexible, generalizable robot control without per-task data collection or model retraining.
Features
- Code-as-policy architecture that writes executable robot programs from natural-language instructions
- Long-horizon planning via reasoning model that decomposes tasks, verifies intermediate outcomes, and re-plans on failure
- Multi-agent coordination where a master agent spawns subagents to control multiple robots concurrently
- Generalist operation across different robot arms, grippers, and camera setups without finetuning or new data collection
- Conversational interface for iterating on robot programs in real time
- Data generation capability, such as autonomously collecting thousands of pick-and-place demonstrations and training an ACT policy from that data overnight