RoboArchitect
A method for automatically designing robot agent harnesses. A coding model iteratively revises control code wrapped around a frozen policy and evolves persistent memory, without updating model weights.
Timeline (last 90 days)
RoboArchitect proposes updating no model weights, letting a coding model automatically design a harness wrapped around a frozen policy and injecting persistent memory into it.
Evaluations across multiple benchmarks and policies show that RoboArchitect raises the average success rate on unseen scenes from 9.7% to 43.3% (4.5x).
RoboArchitect's workflow is that the coding model observes the robot's rollouts, diagnoses the causes of failure, and then iteratively revises the framework wrapping the frozen policy, without updating any model weights throughout the process.
The framework is evolved and optimized over a set of episodes and evaluated on held-out scenes and tasks to avoid overfitting.
In RoboArchitect, memory is not simple conversation history but an experience library that the framework code can read and write, for example, "at a certain grasp pose, the success rate is higher when the gripper depth is 2 cm".
The evolved control strategies often have intuitive physical meanings, such as lifting a stuck robotic arm to free it from an obstacle, or switching to approaching horizontally when the top is blocked.
The biggest gains come when the benchmark is far from the policy's training data: with a general policy as the wrapper, the evolved framework can produce an "expert" that often even surpasses the fine-tuned version.
All Facts (7)
RoboArchitect proposes updating no model weights, letting a coding model automatically design a harness wrapped around a frozen policy and injecting persistent memory into it.
50%UnverifiedEvaluations across multiple benchmarks and policies show that RoboArchitect raises the average success rate on unseen scenes from 9.7% to 43.3% (4.5x).
50%UnverifiedRoboArchitect's workflow is that the coding model observes the robot's rollouts, diagnoses the causes of failure, and then iteratively revises the framework wrapping the frozen policy, without updating any model weights throughout the process.
50%UnverifiedThe framework is evolved and optimized over a set of episodes and evaluated on held-out scenes and tasks to avoid overfitting.
50%UnverifiedIn RoboArchitect, memory is not simple conversation history but an experience library that the framework code can read and write, for example, "at a certain grasp pose, the success rate is higher when the gripper depth is 2 cm".
50%UnverifiedThe evolved control strategies often have intuitive physical meanings, such as lifting a stuck robotic arm to free it from an obstacle, or switching to approaching horizontally when the top is blocked.
50%UnverifiedThe biggest gains come when the benchmark is far from the policy's training data: with a general policy as the wrapper, the evolved framework can produce an "expert" that often even surpasses the fine-tuned version.
50%