The Pulse
Waddle-1 scores 81.4% on Robocurve's real-robot benchmark, with no model-call cap
Waddle Labs' robot-use agent scored 81.4% on Robocurve's real-world task benchmark, ahead of the other listed configurations. The system had no model-call cap, unlike the language models on the leaderboard.

Waddle Labs' Waddle-1 scored 81.4% on a real-world robot task benchmark, the highest mean score listed by evaluator Robocurve. The result is ahead of GPT-6.2 Astra's 65.7%, but the systems were tested under different limits: Waddle-1 had no model-call cap, while language-model configurations were capped at 40 calls.
Waddle Labs introduced the robot-use agent on November 20. Its result on RoboDojo-RC Tier 1 reflects progress across task attempts, rather than the proportion of tasks completed in full.
A cloud agent with local policies
The company says Waddle-1 controls two-armed robots on tasks it has not seen before, without task-specific examples. It calls language models in the cloud to direct lower-level policies that run locally.
Waddle Labs is inviting people interested in deploying robots to request access to a preview API. Robocurve describes Waddle-1 as a full system made up of language models that control the robot through lower-level policies.
In the evaluation, Waddle-1 had one chat per task and a 10-minute limit, with no cap on model calls. Robocurve also applies a limit of 3,600 control steps to vision-language-action models, a separate category represented on the leaderboard.
What the score measures
The benchmark uses six real-world manipulation tasks adapted from RoboDojo, performed on I2RT YAM arms, according to explainx.ai's report on the results. The tasks include capping a pen, sorting objects, stacking bowls and standing up bottles.
The score measures mean progress against task-specific rubrics across 120 attempts per configuration, explainx.ai reports. Those rubrics give credit for stages of progress, rather than recording only whether an attempt was fully successful.
For the bottle task, Robocurve reported that Waddle-1 scored 99% and completed 19 of 20 trials. The evaluator described that task as particularly difficult for language models tested before GPT-6.2 Astra.
Cost and time comparisons
Robocurve estimates Waddle-1's hosted inference cost at $2.46 per task attempt, based on a planned rate of $20 per hour. The leaderboard puts GPT-6.2 Astra's high-effort, ultrafast configuration at an estimated $10.51 per attempt.
The leaderboard lists Waddle-1 at 7.4 minutes per attempt and GPT-6.2 Astra's high-effort, ultrafast configuration at 2.7 minutes. For language models, the time measure is wall-clock time per attempt; those models can end an attempt by indicating that they are done or giving up.
Dream-1 scored 59.5%, with estimated hosted inference cost of $0.15 per attempt and a listed time of 1.5 minutes, according to Robocurve. For vision-language-action models, the time figure is the mean time taken to reach their final score; the evaluator says those models always run to their control-step cap.
Sources
- Original source Robocurve · page 1 robocurve.org
Other pages read
- Waddle Labs x.com
- explainx.ai
- Robocurve · page 2 x.com
In the tracker
Related stories
- TuringSpace test finds spatial action gaps in leading AI models
- Google brings Hydrogen robotics tools to paid Antigravity users
- Humyn Labs raises $47 million to expand robot data operations
- LlamaIndex proposes severity-based scoring for document parsing
- Sakana AI forms security research group focused on AI defense