RoboArchitect Lets Robots Evolve Their Control Harnesses Without Weight Updates

Introduction

A robot’s performance depends on more than the policy that predicts its actions. It also depends on how that policy is used: when to query a vision-language-action model, when to request additional visual information, and when to take direct control of the arm through code. RoboArchitect explores whether this surrounding agent harness can be designed automatically instead of being written by hand for every robot and benchmark.

Optimizing the layer around a frozen policy

RoboArchitect leaves the underlying robot policy unchanged. A coding model observes execution rollouts, identifies why attempts fail, and iteratively revises the control code wrapped around the policy. The harness can decide when to invoke the frozen policy, when to consult vision tools, and when to issue direct commands to the robot.

The system improves its harness on one set of episodes and evaluates it on held-out scenes and tasks. This setup is intended to measure whether the revisions produce reusable control strategies rather than simply memorizing the optimization examples.

Code-based control and persistent experience

The resulting agents often settle on a dynamic division of labor. The learned policy handles actions it already understands, while code-based routines take over for explicit, structured, or recovery-oriented behaviors. The system can discover strategies such as lifting an arm to clear a jam or approaching horizontally when an obstacle prevents a top-down route. These behaviors are not presented as fixed rules supplied in advance; they emerge through the iterative harness-design process.

RoboArchitect also evolves persistent memory. Useful experience from earlier episodes can be stored and used in later decisions. For example, an effective grasp depth for a type of object may become a reference for subsequent attempts. This makes the harness more than a wrapper of control functions: it becomes a mechanism for accumulating experience across episodes.

Results and broader significance

Across multiple benchmarks and policies, average success on unseen scenes rose from 9.7% to 43.3%, while average success on unseen tasks increased from 12.7% to 40.2%. These correspond to roughly 4.5-fold and 3.2-fold improvements, respectively, without updating any model weights. The reported gains are strongest when the evaluation setting differs substantially from the policy’s training distribution. In that situation, a harness built around a general-purpose policy can act like a specialist and sometimes outperform a fine-tuned counterpart.

The work shifts part of the robotics discussion from training larger policies to organizing existing capabilities more effectively. A future robot agent may combine a frozen model, tools, executable control code, and memory rather than relying on a single end-to-end policy. At the same time, the reported results are tied to the benchmarks, policies, and experimental settings in the paper. The reliability of automatically evolved harnesses in more open-ended physical environments remains an important question for further study.

Source: Hugging Face Daily Papers