Safety Activation Steering

• Safety Activation Steering is a framework that adaptively constrains vehicle velocity to maintain safe states under varied steering inputs.
• It uses forward simulation of candidate steering profiles and a receding-horizon optimization to predict and prevent collisions.
• The approach intervenes selectively on acceleration while preserving lateral control, thereby balancing safety with driving comfort.

Safety Activation Steering refers to a class of inference-time interventions that modify a system’s behavior—typically a dynamical system or a neural model—by adaptively constraining, steering, or modulating its action selection to maintain or enhance safety, without fundamentally altering its primary (control or generative) policy. In the context of teleoperated or automated road vehicles, as exemplified by the Steering-Action-Aware Adaptive Cruise Control (ACC) framework, Safety Activation Steering computationally ensures that—even under arbitrary or adversarial steering inputs from an operator—the system never leaves the set of states from which it can be safely stopped without collision. The method generalizes to other domains where robust compliance with safety constraints under input uncertainty or human-in-the-loop operation is essential (Schimpe et al., 2022).

1. Formal Definition of the Safe State Under Steering Uncertainty

Safety Activation Steering in the Steering-Action-Aware ACC introduces a rigorous definition of the safe set, denoted $S$, for the controlled system (teleoperated vehicle):

$$S = \{ z = (x, y, \theta, \delta, v) \mid \forall \ \dot{\delta}(t)\in[-\dot{\delta}_{\rm max}, +\dot{\delta}_{\rm max}], \exists \ a(t)\leq 0 \ \text{s.t.} \ s_{\rm stop}(v) \leq s_{\rm available}(\dot{\delta}(\cdot)) \}$$

where

• $z$ is the full system state (position, heading, steering angle, speed);
• $s_{\rm stop}(v)$ is the minimum stopping distance as a function of current velocity $v$ (with a fixed braking profile, $a_{\rm stop} = -v_{\rm curr}/T_H$);
• $s_{\rm available}(\dot{\delta}(\cdot))$ is the distance that can be traveled without collision, under a particular steering rate profile, while braking.

The state $z$ is considered safe if, under any admissible steering input (including the worst-case possible operator choice), there always exists a braking command that will halt the system before it reaches an obstacle (Schimpe et al., 2022).

2. Forward Simulation of Worst-Case Future Trajectories

Rather than operate conservatively for all possible futures, the ACC samples a finite set of $M$ candidate steering-rate profiles:

$$\dot{\delta}_m = -\dot{\delta}_{\rm max} + \frac{m-1}{M-1} (2\dot{\delta}_{\rm max}), \quad m = 1,\dots,M$$

Each profile is combined with a fixed maximal braking action ($a_{\rm stop}$). Using the (forward-Euler discretized) kinematic bicycle model, the approach simulates rollouts of the full state vector:

$z_{n+1}^{m} = z_{n}^{m} + t_s f(z_{n}^{m}, u_{m}),$

where $u_m = [\dot{\delta}_m, a_{\rm stop}]^\top$, projecting $N$ steps into the future. Collision checking is run along each trajectory to obtain the “available progress” $s_m$ until the first obstacle hit.

3. Risk Quantification and Global Safety Margin

The approach defines $s_{\rm safe} = \min_{m=1,\dots,M} s_m$ as the minimum safe progress over all sampled steering-rate futures: this quantifies the most pessimistic but feasible stopping distance among all admissible steering actions. If $s_{\rm safe} < s_{\rm stop}(v)$ for the current speed, the vehicle is predicted to collide in at least one scenario, triggering safety intervention (Schimpe et al., 2022).

4. Receding-Horizon Safe Longitudinal Optimization

A receding-horizon quadratic program determines the “safe and comfortable” velocity profile $v_n$ that ensures all safety constraints are respected at every future step:

• Discrete motion model (position/speed/acceleration/jerk coupling)
• Progress constraint: $s_{n+1}\leq s_{\rm safe}$
• Lateral acceleration constraint: $| \kappa_{n+1} v_{n+1}^2 | \leq a_{\rm lat, max}$
• Smoothness (jerk, acceleration bounds) via soft constraints/slack variables

The optimization minimizes deviation from the operator’s velocity command while enforcing that the commanded profile always remains executable without violating any safety constraint under possible steering (Schimpe et al., 2022).

5. Override Logic: Selective Intervention in the Control Loop

If, at any moment, the optimized $v_\text{safe}$ drops below the human operator’s velocity request $v_\text{des}$, the ACC overrides only the longitudinal command—throttling or braking to maintain $z$ within the safely-stoppable set $S$. In contrast to a full model predictive override (which might also alter steering), Steering Action-Aware ACC intervenes exclusively in the longitudinal channel, leaving lateral authority (steering) under operator control (Schimpe et al., 2022). This selective override is driven by the shrinking of $s_{\rm safe}$ as predicted obstacle-induced hazards become imminent.

6. Comparative Performance and Safety-Comfort Trade-off

In direct simulation against a baseline model-predictive controller that can override both speed and steering, Steering Action-Aware ACC demonstrates:

Criterion	Steering ACC	Full MPC Baseline
Lateral Freedom	Unchanged (driver has full control)	Artificially limited
Accident Risk	No collisions	No collisions
Speed Retention	Near nominal unless stopping needed	Slightly more conservative
Comfort (Jerk)	Comparable to MPC	Comparable

Empirically, the system brakes only when a hazard is genuinely imminent; decelerations and stops are well-timed to remain within safe stopping bounds, and lateral maneuvers not implicated in safety remain under operator control (Schimpe et al., 2022).

7. Physical Demonstration and Practical Robustness

Experiments conducted on a 1:10 scale teleoperated vehicle (live perception via stereo camera and lidar, realistic control latency and perception noise) confirm that the Safety Activation Steering strategy persistently maintains vehicle state inside the safe set $S$. Notably, when an operator steered into a region that would result in collision if unmitigated, the system automatically reduced $v_\text{safe}$ to zero, preventing any forward motion until a safe trajectory became available (Schimpe et al., 2022).

This protocol has proven robust in real-world latency and sensor-imperfect conditions, validating that it suffices to intervene only on velocity commands to enforce robust collision avoidance under arbitrary steering.

---

References

• “Steering Action-aware Adaptive Cruise Control for Teleoperated Driving” (Schimpe et al., 2022)