Higher-Order ADAS Coordination

• Higher-order ADAS is a hierarchical control system that coordinates modules like ACC, AEB, and LKS under a unified decision-making framework.
• It employs adversarial imitation learning and derivative-free optimization to fuse sensor data from 360° LIDAR for real-time, robust control.
• Evaluations in simulated multi-lane highway environments show near-expert performance, enhancing both vehicle safety and efficiency.

Higher-order advanced driver assistance systems (ADAS) enhance the autonomy and safety of vehicles by enabling the simultaneous coordination of multiple foundational ADAS functions—such as adaptive cruise control (ACC), emergency braking (AEB), and lane-keeping/lane-change assist (LKS)—under a unified high-level decision-making framework. This hierarchical approach is essential for autonomous driving systems tasked with operating in complex, multi-agent environments, where effective arbitration among several ADAS modules is necessary for nuanced, real-time control and safety assurance. A prominent instantiation utilizes a policy trained via adversarial imitation learning, which directly gates low-level ADAS modules according to high-level situational assessment derived from sensor data, most notably 360° LIDAR arrays, ensuring robust operation in multi-lane highway scenarios (Shin et al., 2019).

1. Problem Framing and System Architecture

Higher-order ADAS is modeled as a partially observable Markov decision process (POMDP) with a well-defined observation and action space. The observation space $O \subset \mathbb{R}^n$ comprises raw and derived LIDAR signals encapsulating spatial and kinematic states: ranges $d_i \in [0, r_{\text{max}}]$ for each LIDAR beam, and relative speeds $v_i$, forming the vector $$o = (d_1, \ldots, d_N, v_1, \ldots, v_N)^\top \in \mathbb{R}^{2N}$$. The action space $A = \{1, \ldots, 5\}$ represents discrete maneuver classes: maintain, accelerate, decelerate, lane-left, and lane-right.

These five high-level action primitives map directly to underlying ADAS modules: ACC is engaged by "accelerate," a combination of ACC and AEB is activated by "decelerate," and LKS is responsible for lane-centric actions. This mapping constrains the gating policy such that exactly one, or a specified composition, of ADAS modules is active at any decision epoch, reflecting a hard-hierarchical supervisory control scheme.

2. Adversarial Imitation Learning as the Supervisory Mechanism

Coordination of multiple ADAS modules is learned through a randomized adversarial imitation learning (RAIL) framework, an extension of generative adversarial imitation learning (GAIL). In this setting, an expert policy $\pi_E$ (typically constructed via reinforcement learning with hand-crafted logic) generates demonstration trajectories, which are used to supervise the learning of the ADAS coordinator policy $\pi_\theta$.

The core objective is

$$\min_{\pi_\theta} \max_{D_\phi} \mathbb{E}_{\pi_\theta} [\log D_\phi(s,a)] + \mathbb{E}_{\pi_E}[\log(1 - D_\phi(s,a))] - \lambda H(\pi_\theta)$$

where $D_\phi$ is a discriminator parameterized by $\phi$ and $H(\pi)$ is the entropy regularization term to encourage exploration. RAIL substitutes the usual cross-entropy GAN objective with a least-squares GAN (LS-GAN) loss:

$$L_{LS}(D) = \frac{1}{2} \mathbb{E}_{\pi_E}[(D(s,a)-1)^2] + \frac{1}{2} \mathbb{E}_{\pi_\theta}[(D(s,a))^2]$$

The policy's reward signal is defined via the logit transform $r(s,a) = \log D_\phi(s,a) - \log(1 - D_\phi(s,a))$, and the ultimate policy objective is to maximize $\mathbb{E}_{(s,a) \sim \pi_\theta}[r(s,a)]$.

3. Derivative-Free Parameter Optimization and Policy Networks

Policy and discriminator are both implemented as shallow multilayer perceptrons (MLPs). The policy network $\pi_\theta$ uses two sets of weights ($\theta^i \in \mathbb{R}^{n \times h}$ and $\theta^o \in \mathbb{R}^{h \times p}$) and a nonlinear activation (e.g., ReLU or $\tanh$):

• Hidden state: $h = \sigma(\theta^i s_\text{norm})$
• Output: $\pi(s) = \text{softmax}(\theta^o h)$

The discriminator $D_\phi(s,a)$ outputs a scalar in $[0,1]$.

RAIL leverages derivative-free optimization via an adaptation of Augmented Random Search (ARS):

• $N$ independent Gaussian perturbations $\delta_k$ are sampled for the policy parameters.
• Policies $\theta_\pm = \theta_t \pm \nu \delta_k$ are rolled out; segment reward differences $\Delta r_k = r^+_k - r^-_k$ are computed.
• Parameters are updated by

$$\theta_{t+1} = \theta_t + \frac{\alpha}{N \sigma_R} \sum_{k=1}^N \Delta r_k \delta_k$$

where $\sigma_R$ is the empirical standard deviation of the batch's rollout rewards.

Policies are initialized by behavioral cloning (BC) from the expert trajectories to warm-start imitation and stabilize adversarial training.

4. Sensor Preprocessing and Multi-Modal Integration

Raw sensor input consists of 24 LIDAR beams spaced at 15° increments, delivering both distance and relative velocity per spatial direction. The system constructs a $2N$-dimensional observation vector $s = (d_1, \ldots, d_N, v_1, \ldots, v_N)$. Online normalization is employed by maintaining running estimates of the mean $\mu_t$ and covariance $\Sigma_t$, with normalized states $s_\text{norm} = \Sigma_t^{-1/2}(s - \mu_t)$ provided to the policy MLP. This approach supports real-time integration of rich sensor modalities required for robust multi-lane navigation and dense traffic contexts.

5. Hierarchical Module Gating and Action Arbitration

At each timestep, the higher-order ADAS policy (the "supervisor") selects an action $a_t \in \{1, \ldots, 5\}$. A one-hot vector $g \in \{0,1\}^3$ determines active low-level controllers:

• $$g_\text{ACC} = I[a_t \in \{\text{accelerate},\text{decelerate}\}]$$
• $g_\text{AEB} = I[a_t = \text{decelerate}]$
• $$g_\text{LKS} = I[a_t \in \{\text{lane-left},\text{lane-right}\}]$$

Only one ADAS function—or a deterministic combination—executes for the given high-level decision. This discrete, hard-hierarchy gating is central to the RAIL approach, though the architecture is amenable to soft, continuous mixing weights $w(s) \in \Delta^M$ should extension to blended control be sought.

6. Empirical Evaluation and Performance Analysis

The efficacy of higher-order ADAS coordination via RAIL is established in a simulated five-lane highway environment with randomly spawned vehicles and stochastic, but non-colliding, agent behavior. Key evaluation metrics include average speed, frequency of overtakes and lane-changes, longitudinal rewards (speed-centric), and lateral rewards (maneuver decisiveness). Table 1 summarizes representative performance over 16 episodes (40 trajectories):

Metric	RAIL (2-layer)	RAIL (1-layer)	Expert
Speed (km/h)	70.38	65.00	68.83
Overtakes	45.04	40.03	44.48
Lane-changes	15.01	13.05	14.04
Longitudinal	2719.38	2495.57	2642.11
Lateral	–122.98	–175.60	–132.52

The 2-layer RAIL policy matches or slightly surpasses the expert in speed and overtaking, while maintaining similar lane-changing patterns. Even linear (1-layer) policies reach approximately 90% of expert performance. Sample efficiency is superior: near-expert performance is achieved with only a few dozen expert trajectories, outperforming GAIL+TRPO/PPO baselines in both stability and data efficiency. This indicates that high-level, adversarially-trained, derivative-free parameter optimization techniques are effective for real-world ADAS module coordination, particularly when rich sensor streams are integrated and strict real-time constraints must be satisfied (Shin et al., 2019).