Artificial intelligence
Rapid motor adaptation
Definition
Rapid motor adaptation is a learned control method in which a robot infers changing environmental or physical conditions from a short history of onboard observations and adjusts its policy online. The term is associated with an architecture that pairs a base locomotion policy with an adaptation module.
Also known as: RMA
Updated
Infer hidden conditions from recent motion
Rapid motor adaptation uses recent sensor and action history to infer a compact representation of conditions that the robot cannot measure directly. These conditions can include terrain properties, payload, actuator behaviour, and contact response. A base policy conditions its actions on that representation, so the robot can change its behaviour without retraining its weights during deployment.
The RMA paper describes two learned components. A base policy is trained in simulation with privileged information about the simulated environment. An adaptation module then learns to predict the base policy's environmental representation from a history of proprioceptive observations and actions. At deployment, the adaptation module replaces the unavailable privileged inputs.
A learned form of online adjustment
RMA is related to domain randomization because varied simulated conditions teach the base policy how behaviour should change. It is also related to adaptive control, but the adaptation signal is a learned latent representation rather than necessarily an identified set of physical parameters.
The authors deployed their policy on an A1 quadruped without real-world fine-tuning and reported locomotion across several surfaces and payload conditions. That is evidence for the particular trained system and experimental range, not a claim that any robot with an adaptation module will handle arbitrary changes.
Fast inference is not unlimited adaptation
The module can only infer distinctions represented in its training data and observation history. Different physical causes may produce similar short-term sensor traces, while a sudden failure or terrain transition may occur faster than the history window can reveal it. Errors in the learned environmental representation can then lead the base policy toward an unsuitable action.
RMA also does not remove the sim-to-real transfer problem. It changes how a policy responds to variation, but the simulator, randomisation ranges, sensors, action timing, and safety limits still determine what reaches the real robot safely. Results on one quadruped do not establish the same adaptation speed or stability for a humanoid with different contacts and dynamics.
Sources
Related terms
Domain randomization
Domain randomization varies properties of training environments to encourage a learned model or policy to work across changing conditions. In robotics it often randomizes simulated appearance, physical parameters, or both to support transfer to real hardware.
Sim-to-real transfer
Sim-to-real transfer applies a model, policy, or behavior developed in simulation to a physical system. Its central challenge is the difference between the simulated environment and the robot, sensors, and interactions encountered in reality.
Adaptive control
Adaptive control is a control approach that changes selected controller parameters online to compensate for uncertain or changing system dynamics. In robotics, it can maintain tracking as payload, inertia, friction, or other modeled quantities differ from their assumed values.
Proprioception
Proprioception in robotics is sensing the robot's own motion, configuration, and internal physical state. Typical proprioceptive inputs include joint encoders, inertial measurements, and signals associated with actuator effort or contact.