Does knowledge about perceptual uncertainty help an automated driving agent?

Uncertainty estimates rarely reach the decision maker

Perception uncertainty is usually measured inside the perception module, for example through false positive detection rates or calibration errors on camera images. What that uncertainty should mean for the next driving action is a separate question, and the paper targets exactly that interface. Its research question is whether a reinforcement learning agent learns to adapt its behavior when it is told that its perception is currently uncertain.

The setup separates two things that are easy to conflate: the perception being wrong, and the agent knowing that it is wrong. The authors perturb the agent's view in a controlled way and, in some scenarios, add a signal describing the perturbation. That lets them compare behavior under unreliable input with behavior under unreliable input plus knowledge about the unreliability.

Perturbing vehicles in a bird's-eye segmentation mask

The agent sees the road through a colorized semantic segmentation mask in bird's-eye view, flattened and gray-scaled into a 100-dimensional vision component. A separate 6-dimensional non-visual component carries throttle and braking intensity, velocity, normalized velocity, orientation and distance to the lane center. The uncertainty component is a 4-dimensional one-hot vector, or an empty singleton set when no uncertainty information is given.

Uncertainty is created by removing vehicles from the ego-vehicle lane in the mask. Four perturbation cases exist: the closest front vehicle is overlooked (VEXV), the following vehicle is invisible (XEVV), both front vehicles are invisible (VEXX), and all vehicles except the ego-vehicle are invisible (XEXX). A fifth case, VEVV, shows all vehicles and counts as correct perception. Because the perturbation is constructed rather than estimated, the uncertainty supplied to the agent is exactly quantifiable, which is what makes the comparison controlled.

Four scenarios cross perception quality with knowledge

The experiments combine two binary factors into four scenarios: correct or perturbed perception, and uncertainty information absent or present. Scenario 1 is correct perception without information, scenario 2 is perturbed perception without information, scenario 3 is perturbed perception with information, and scenario 4 is correct perception with information.

The agent is trained with proximal policy optimization for two million time steps per experiment, with the best model selected on validation episodes and then tested under the perceptual perturbations. The action space is a single continuous value between -1 and 1, where non-negative values accelerate and negative values brake, with an inertia model damping the applied action. Steering is excluded; the task is to drive a straight 150-meter route as fast as possible without colliding with other road users.

A reward that weights speed against elapsed time

The reward is not a plain velocity term. It multiplies the agent's momentary velocity toward the end point by a factor that decays linearly with time, so speed achieved earlier in an episode counts for more. A constant term still rewards momentary speed regardless of time, to allow for situations where the agent must slow down.

Failures return a negative reward when the agent collides, exceeds the maximum time limit, or does not start driving soon enough; finishing successfully returns a positive reward. This design matters for interpreting the results: an agent that simply waits for other vehicles to clear the route would lose reward, so the comparison between scenarios reflects a genuine trade-off between pace and collision avoidance rather than an artifact of a speed-only objective.

Perturbed perception alone produces defensive driving

The central behavioral finding is that an unreliable observation space, modeled by perturbed perception, leads the agent to drive defensively. Trained and tested on perturbed perception without uncertainty information, the agent generally acts cautiously to avoid collisions. The paper reports that the agent is able to finish its task in all three training scenarios when tested on the training scenario.

A sharper contrast comes from mismatch. When the agent is trained with correct perception but evaluated with perturbed perception, it fails consistently. That result locates the value of the perturbation experiments: an agent tuned to clean input does not simply slow down when input degrades, it breaks down. Defensive behavior in the perturbed-training condition is therefore a learned response to unreliable input, not a generic fallback.

The uncertainty signal restores pace without dropping caution

Adding the uncertainty information directly to the observation space changes the behavior. The agent trained on perturbed perception with uncertainty information adapts to the specific situation: it acts more defensively under a perturbed perception, and it increases its pace when the uncertainty information states that perception is correct. In general, this agent accomplishes its task faster while at the same time accounting for risks.

The comparison that carries the paper's answer is therefore between scenario 2 and scenario 3, both trained on perturbed perception. The difference is not whether the agent is cautious but whether caution is conditional. Without the signal, defensive driving applies broadly; with it, the agent can distinguish a perturbed frame from a correct one and spend caution only where the perception warrants it.

What the braking-only straight route does not cover

The evidence comes from a proxy task in the CARLA simulator: a straight 150-meter route, dynamic objects, and a two-action continuous control space with no steering. The uncertainty is artificial and perfectly labeled, since the perturbation type is known by construction. Real perception uncertainty would have to be estimated, and estimation errors would enter the observation rather than a clean one-hot vector.

These conditions bound the claim. The paper shows that information about perceptual uncertainty helps an agent in this controlled driving task; it does not show that a specific uncertainty estimator would deliver the same benefit on real sensor data, nor that the behavior transfers to steering, intersections, or more complex routes. The reported outcomes are also behavioral measures within episodes, such as failure, traveled distance, used time steps, distance to the front vehicle and brake frequency, rather than a safety guarantee.

Conditional caution is the result worth carrying forward

The paper's most useful distinction is not that uncertainty information makes an agent faster or safer, but that it makes caution conditional on the current perturbation type. An agent trained on unreliable perception without that signal pays a pace penalty everywhere; an agent given the signal keeps defensive behavior where perception is perturbed and recovers pace where it is not. For anyone evaluating uncertainty estimation for driving, that reframes the downstream test: the question is whether the estimate lets the planner separate situations, not merely whether the estimate is well calibrated in the perception domain.

The scenario split also sets the priority for follow-up work. The consistent failure of the agent trained on correct perception and tested on perturbed perception indicates that robustness to degraded input is not free, and that training conditions must match the deployment conditions an agent will face. The uncertainty channel is what makes that matching cheaper in pace terms within this task.

Sources used in this answer