Hardware & control
Safety filter
A safety filter is a supervisory component that evaluates a proposed robot action, control input, plan, or reference against declared constraints and passes it through, modifies it, or substitutes a safer alternative before execution. It separates nominal task behaviour from constraint enforcement, but it only protects against hazards represented by its model, observations, and safety specification.
Also known as: runtime safety filter, safety shield, runtime shield
Updated
The nominal controller proposes; the filter supervises
A safety filter sits between a source of behaviour and the controlled system. The source may be a learned policy, motion planner, teleoperator, or generative model. The filter evaluates the proposal using the current state and a safety specification, then passes compliant behaviour through or intervenes before execution.
The predictive safety-filter formulation checks whether a proposed control input can satisfy future state and input constraints. If not, it supplies an alternative input while trying to modify the original as little as necessary. Other filters use reachability, control barrier functions, rule-based shields, or combinations of learned and model-based components.
Filters can act on actions, trajectories, or references
The protected object is part of the definition. A low-level filter may adjust the actuator command at every control step. A predictive filter may assess a short future plan. A motion-generation filter may edit an entire reference before a separate whole-body controller tracks it.
CSF operates at two points. During generation, it filters the motion estimate at each sampling step. During execution, its runtime shield can preserve an already executed prefix and replace the uncommitted remainder when a protected entity enters the scene.
Minimal intervention is a design goal, not automatic proof
Many safety filters optimise for the closest allowed alternative so useful task behaviour survives. That trade-off depends on the distance measure and on which constraints are active. A conservative filter may stop too often; a permissive or poorly observed filter may miss a hazard.
The word “safety” does not imply certification. A filter cannot enforce hazards it does not model or perceive, and a valid filtered command may still be executed incorrectly by the physical system. Evaluation should report false interventions, missed hazards, timing, feasibility failures, task degradation, and hardware outcomes—not only whether an optimisation returned a solution.
Dataset records should keep the nominal proposal, safety specification, state estimate, active constraints, filtered output, reason for intervention, latency, and measured execution separately. Otherwise downstream users cannot distinguish naturally safe behaviour from behaviour made safe only by an external supervisor.
Sources
Related terms
Hardware & control
Control barrier function
A control barrier function is a scalar function used to describe a safe set of system states and constrain control choices so trajectories remain inside that set. It is commonly enforced through an optimisation problem that modifies a nominal control input only when needed; its guarantee depends on the dynamics model, state estimate, constraints, and assumptions used to define the safe set.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.
Hardware & control
Feedback control
Feedback control is a closed-loop control method that measures a system’s current output or state, compares it with a target and adjusts the command using the resulting error. In robotics, feedback can correct joint, end-effector, balance or force errors as new sensor measurements arrive.
Hardware & control
State estimation
State estimation is the process of inferring variables that are not known exactly—such as a robot’s base pose, velocity, contact state or sensor bias—from noisy measurements, control inputs and a model of how the system evolves. An estimator should provide both an estimate and, where possible, a representation of its uncertainty.
Hardware & control
Model predictive control
Model predictive control is a feedback-control method that uses a dynamics model to optimise a sequence of future actions, applies the first action or short part of that sequence, and replans from an updated state estimate. In receding-horizon MPC, the planning window moves forward at each update.