Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 2 min read

Waddle-1 scores 81.4% on RoboCurve's RoboDojo-RC Tier 1 robot benchmark

Waddle-1 at a glance RoboCurve's RoboDojo-RC Tier 1 board, as of November 20, 2026, lists Waddle-1, Waddle Labs' robot-control agent system, first with a mean score of 81.4% over 120 task attempts, ahead of GPT-6.2 Ast

Waddle-1 at a glance

RoboCurve's RoboDojo-RC Tier 1 board, as of November 20, 2026, lists Waddle-1, Waddle Labs' robot-control agent system, first with a mean score of 81.4% over 120 task attempts, ahead of GPT-6.2 Astra at 65.7%. Waddle-1 ran as a full system with no call cap, unlike the capped model configurations below it.

What changed with Waddle-1

RoboCurve published Waddle-1's result on its RoboDojo-RC Tier 1 board on November 20, 2026: a mean score of 81.4% across 120 attempts at real-world robot tasks, the highest score on the board. Waddle-1 is the robot-use agent system from Waddle Labs, an applied AI lab whose agents control robots by writing code: they break a goal into subtasks, view camera feeds, write control code and call lower-level action models, then hand back a program you can run and revise by talking to the agents โ€” a workflow Y Combinator describes as 'the Claude Code for robotics'.

The rest of the board shows the gap. As of November 20, 2026, the strongest standard model configuration, GPT-6.2 Astra (high), scored 65.7%, followed by Dream Labs' Dream-1 at 59.5%, Claude Fable 5.2 (high) at 57.5% and Gemini 4 Carbon (high) at 54.5%, while NVIDIA's GR00T N1.7, a purpose-built robot policy, scored 9.5%. RoboCurve runs every configuration over the same 120 attempts at tasks such as capping a pen, plugging a charger into a power strip and pouring a drink.

The comparison carries a material caveat, stated by RoboCurve on the board itself: the model configurations are capped at 10 model calls per attempt (robot policies get 3,600 control steps), while Waddle-1 ran as a full system โ€” one chat per task, a 10-minute limit and no call cap โ€” and averaged 7.4 minutes per attempt. RoboCurve estimates its cost at $2.46 per attempt, from a planned $20-per-hour hosted price; Waddle Labs has not published pricing. Waddle-1 is not openly available โ€” access is by request on Waddle Labs' site โ€” and so far the score rests on this one board alone.

Waddle-1 in numbers

Values are copied from the linked source on the date shown. They are not IA Compa measurements.

  • RoboDojo-RC Tier 1 mean score (RoboCurve), as of Nov 20, 2026: 81.4%

Why Waddle-1 matters for businesses

Why it matters: if agent-written robot control holds up beyond a single board, a small manufacturer or warehouse could put a robot on a new physical task by describing it and reviewing the program the agent writes, instead of hiring integrators or collecting training data for every machine and task. Waddle-1's result is an early signal gathered under a friendlier run regime than the models below it faced, but it points at the shift coding agents already brought to software: the model writes the code, a person supervises, and the cost of trying something new falls.

Sources

Last verified: November 21, 2026. IA Compa summarizes public announcements for business readers. We are not affiliated with the companies in these stories.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.