Skip to main content
Back

Robocurve Finds GPT-6 Astra Controlling Robot Arms Without Specialized Training

A new study from Robocurve measures how fast general-purpose language models are learning to drive robot arms, and the curve is steep.

  • Robocurve's Christopher Leet et al. systematically evaluate nine frontier LMs on six physical manipulation tasks via arXiv 2610.25287.
  • Mean task score roughly doubles every 53.9 days between July and October 2026.
  • Best subtask completion rate improved 4.4x; lowest cost per subtask fell 6.2x in the same window.
  • No model wins on score, speed, and cost at once; Opus 5.5 offers the best balance.
  • Failures come from action-budget exhaustion or unrecoverable errors, not model refusals.
  • Evaluation harness Inspect Robots is open-source and supports 260+ LLMs plus 40 VLAs.

A team led by Christopher Leet at Robocurve, an independent public benefit corporation that evaluates frontier AI on physical hardware, has published a new study of general-purpose multimodal models controlling a real robotic arm. The models rarely completed manipulation tasks within fixed action limits, but their benchmark scores rose sharply between July and October 2026. The result suggests that robotics teams can already prototype with generalist models and should track how quickly those systems are approaching specialized robot policies.

The title, The Bitter Lesson Comes to Robotics, refers to Rich Sutton’s observation that general learning methods powered by increasing computation have repeatedly overtaken hand-engineered approaches. Robotics has remained more dependent on vision-language-action models, or VLAs, trained specifically on robot demonstrations. Robocurve tested whether general-purpose models can narrow that advantage without robot-specific fine-tuning.

Camera frames in, arm poses out

The headline comparison covered GPT-6 Astra, Opus 5.5, and Opus 5 on six real-world tasks adapted from RoboDojo. The full paper expands the evaluation to nine frontier models across several reasoning-effort and service-tier settings.

Element Reported setup
Hardware I2RT YAM bimanual arms
Task suite Six adapted RoboDojo tasks, named RoboDojo-RC Tier 1
Model input Camera frames and task instructions
Model output End-effector poses issued through a tool interface
Fine-tuning No robot-specific training
Variables Model, reasoning effort, and service tier

An end-effector pose specifies the target position and orientation of a gripper. This interface lets a language model direct the arm through structured tool calls while the robot stack handles lower-level motion. Because Robocurve adapted the tasks and assigned the suite a new name, its scores describe RoboDojo-RC Tier 1 and require aligned protocols for direct comparison with other RoboDojo results.