Path-Consistent Braking in Safety Systems

• Path-consistent braking is a safety strategy that modulates the timing of a robot's trajectory without altering its geometric path.
• It integrates braking laws, modal decoupling, and real-time reachability analysis to ensure the system stays within safe operational bounds.
• This approach enables robust deployment of learning-based policies by preserving nominal motion and preventing unsafe lateral deviations.

Path-consistent braking refers to a class of safety intervention strategies in robotic and autonomous systems in which the temporal evolution (velocity, acceleration) of a trajectory is modulated—typically slowed or stopped—while strictly maintaining the robot’s geometric path as prescribed by a nominal policy or reference plan. Within this paradigm, safety filtering or emergency maneuvers act exclusively by altering the timing along the executed path, never deviating laterally or introducing novel motion modes that could displace the system into out-of-distribution states relative to its training data or known safe manifolds. Path-consistent braking has emerged as a robust and theoretically grounded approach for guaranteeing safety under dynamic constraints, real-time requirements, and formal policy consistency demands.

1. Formal Definitions and Conceptual Foundations

A safety intervention is path-consistent if, when applied to an intended trajectory $\hat x(t)$ generated by a policy (e.g., a diffusion policy, model-predictive controller, or feedback planner), the system’s true state $x(t)$ remains within a tight neighborhood of this reference for all times—typically, $\|x(t) - \hat x(t)\| \leq \epsilon$ for some small $\epsilon$ determined by tracking error bounds or model tolerances. Crucially, only the trajectory parameterization in time is altered: velocity and acceleration along the path may be reduced or driven to zero as needed for safety, but the path geometry is never “deflected” or reassigned (Römer et al., 9 Nov 2025). This property ensures that the robot remains within the distribution of training data for learning-based policies, preventing unpredictable or unsafe behavior that would result from lateral deviations.

2. Canonical Models and Control Architectures

The core of path-consistent braking is the coupled design of system dynamics modeling, control input computation for braking, and real-time reachability analysis.

• Robot Joint Space: For manipulator-type systems, the state is $x = [q;\,\dot q] \in \mathbb{R}^{2n}$, with positions $q$ and velocities $\dot q$, and dynamics $\dot x = f(x,u) = [\dot q;\;u]$ for acceleration or torque control inputs $u \in \mathbb{R}^n$ (Römer et al., 9 Nov 2025, Hamad et al., 2023).
• Braking Input Computation: The standard braking law uses

$$u_{\text{brake}}(t) = -\text{sign}(\dot q(t)) \cdot a_\text{max}$$

where $x(t)$0 is the per-joint deceleration limit. More general approaches formulate braking as a quadratic program enforcing acceleration, jerk, and actuation constraints for smoother stopping (Römer et al., 9 Nov 2025, Hamad et al., 2023).

• Bang-Bang Control in Nonholonomic Systems: For mobile robots with discrete brake actuators, control is organized around logical “actions”—turn, go straight, or stop—where each action corresponds to a fixed wheel brake state. State space is partitioned into regions triggering each action to converge the platform onto the path with prescribed error dynamics (Divan et al., 2016).

3. Real-Time Reachability and Safety Filtering

Path-consistent braking leverages set-based forward reachability to guarantee safety:

• Reachable Set Under Braking: For an admissible braking input set $x(t)$1 (such as $x(t)$2), the reachable set $x(t)$3 contains all possible states the robot could occupy under worst-case braking over a horizon $x(t)$4. Fast computational techniques such as zonotopes or interval propagation enable these checks in real time (Römer et al., 9 Nov 2025).
• Obstacle Occupancy Prediction: Potential dynamic obstacles $x(t)$5 are predicted in state space. Safety holds if

$x(t)$6

before continuing along the nominal path; otherwise, braking is triggered.

• Safety Filter Algorithm: At every control period $x(t)$7, measure state, update obstacle prediction, compute reachable sets, and switch control to $x(t)$8 if intersection is detected (Römer et al., 9 Nov 2025, Hamad et al., 2023).

4. Methodological Variants: Modal Decoupling, Conservative and Optimized Braking

Several implementation strategies exist depending on the system and computational requirements:

• Modal-Space Transformation: By linearizing the joint-space inertia matrix $x(t)$9 and using its eigendecomposition $\|x(t) - \hat x(t)\| \leq \epsilon$0 at the initial braking state $\|x(t) - \hat x(t)\| \leq \epsilon$1, the system decouples into independent modal coordinates, allowing one-dimensional time-optimal (bang–bang) braking for each mode (Hamad et al., 2023).
• Conservative (Hyperrectangle) Scaling: A rapidly computable, conservative assignment of actuator limits in modal coordinates ensures feasibility in the original joint space and is suitable for real-time safety filtering at rates $\|x(t) - \hat x(t)\| \leq \epsilon$21 ms/cycle (Hamad et al., 2023).
• Offline-Optimized Scaling: Precomputed maximal feasible modal limits (via a static optimization) yield less conservative but more computationally intensive stopping trajectories, used when slightly longer evaluation time is acceptable (Hamad et al., 2023).
• Backup Set and Barrier-Based Braking for Vehicles: In ground vehicles on variable friction, backup control barrier functions define admissible side-slip/yaw regions and a backup controller, ensuring path-consistent braking without exceeding lateral motion safety envelopes (Gacsi et al., 17 Oct 2025).

5. Comparative Evaluation: Task Success, Path Consistency, and Safety

Empirical studies across application domains highlight significant advantages of path-consistent braking:

Method	Task/Stop Success Rate	Path Consistency (Deviation)	Computation Time
Path-consistent Braking	97% (Safe-Success)	$\|x(t) - \hat x(t)\| \leq \epsilon$35 mm (manipulator)	$\|x(t) - \hat x(t)\| \leq \epsilon$4–$\|x(t) - \hat x(t)\| \leq \epsilon$5 ms
Reactive CBF Filtering	0% (Safe-Success)	Out-of-distribution states	Varies
Max-brake Only	Shortest stop, unsafe	Spins out (vehicles)	$\|x(t) - \hat x(t)\| \leq \epsilon$61 ms

• In manipulation (e.g., human-robot handover), path-consistent braking (PACS) achieves a 97% safe-success rate with only a 3% drop in overall task success, while CBF-based filters yield 0% because of excessive deviation from learned behaviors (Römer et al., 9 Nov 2025).
• In joint-space braking for arms, the maximal lateral deviation from the geometric path remains within 5 mm, well under ISO safety zone tolerances (Hamad et al., 2023).
• In vehicle braking, naive maximum-braking violates lateral safety limits (spinout), while path-consistent, backup-CBF QP approaches maintain safe side-slip and yaw without excessive stopping distance increase (Gacsi et al., 17 Oct 2025).

6. Connections to Related Paradigms and Limitations

Path-consistent braking differs fundamentally from reactive safety interventions such as control barrier functions (CBFs) and general trajectory replanning:

• Reactive Strategies (CBFs): Modify the robot’s control input to enforce pointwise safety (barrier invariance) but frequently deflect the robot laterally, entering states never covered in policy or demonstration data, resulting in degraded task performance and possible loss of stability (Römer et al., 9 Nov 2025, Gacsi et al., 17 Oct 2025).
• Hybrid Feedback and Bang–Bang Approaches: For nonholonomic or underactuated systems, hybrid strategies using logical state partitioning, discrete actions, and Lyapunov analysis can realize stable, asymptotically convergent path tracking and braking without fine-grained force or torque control (Divan et al., 2016).
• Constraints and Open Questions: Although path-consistent braking ensures tight geometric adherence, its effectiveness is dependent on accurate forward models (for reachability), precise actuator constraints, and sufficiently rich policy training data. Overly conservative braking can increase completion time; underconservative models may result in insufficient safety margins.

7. Representative Applications and Practical Deployment

• Safe Deployment of Learning-based Policies: PACS enables deployment of diffusion policies in human-robot interaction by preventing unsafe, out-of-distribution behaviors arising from safety overrides (Römer et al., 9 Nov 2025).
• Robotic Manipulator Emergency Stopping: Modal-space decoupling strategies yield real-time, predictable stopping trajectories suitable for collaborative and industrial robots, with strict adherence to certified safety zones (Hamad et al., 2023).
• Automotive Braking on Asymmetric Surfaces: Backup set-based path-consistent braking frameworks ensure that driver-assist or autonomous braking does not lead to lateral instability, even under torque and friction constraints (Gacsi et al., 17 Oct 2025).
• Assistive Mobile Robots: Bang–bang brake approaches provide stable and comfortable path following for wheeled platforms with minimal sensor and actuation requirements (Divan et al., 2016).

A plausible implication is that path-consistent braking frameworks generalize across a wide range of embodied systems, under the unifying principle of modifying timing, not geometry, for safety. This paradigm simultaneously preserves policy success rates and provides rigorous, real-time safety guarantees.