Actuator

An actuator is a device that converts energy (usually electrical energy) into mechanical motion; it is the core component of a robot joint, usually composed of a motor, reducer, encoder, and driver.

The executor refers to the underlying model that carries the agent's operation, and the harness is the execution framework wrapped around the model, including prompt templates, tool-calling protocols, loop control, result parsing, etc.

Existing recursive self-improvement methods fall into two categories: one freezes the executor and only optimizes the tool harness, which prevents the agent from learning new actions; the other directly fine-tunes the executor on failed trajectories without attribution.

In the context of AdaRSI, the executor refers to the neural network body that carries the action generation capability (usually the weights of a VLA model), while the tool harness is the scheduling and orchestration layer wrapped around the executor.

Modifying the tool harness only requires adjusting code or prompts without gradient updates; fine-tuning the executor requires collecting demonstration data, backpropagation and parameter updates, which is costly and may introduce catastrophic forgetting.

Proposals should be stored in a location writable by the executor, and approvals should be stored in a path to which the executor cannot push.

Responsibilities should be separated between the Agent (responsible for proposing, with no sending or credential modification) and the Executor (responsible for acting, only running proposals marked as approved and atomically checking status).

The number of high-performance actuator manufacturers capable of stable supply globally can be counted on one hand.

A humanoid robot typically has dozens of joints; if the yield rate of a single actuator is 99%, the overall unit yield rate will drop to approximately 74% after stacking 30 joints.