What Probing Reveals about Autonomous Driving: Linking Internal Prediction Errors to Ego Planning

Technical Summary: What Probing Reveals about Autonomous Driving: Linking Internal Prediction Errors to Ego Planning

Problem Statement
Current autonomous driving policies, trained on large-scale datasets and evaluated in closed-loop simulators, often exhibit strong performance in nominal scenarios. However, this high-level success can mask flawed reasoning, unsafe heuristics, and a lack of true understanding of surrounding vehicle dynamics. Standard closed-loop metrics (e.g., collision rates, goal progress) do not reveal whether a policy genuinely predicts the motion of other agents or plans safe trajectories based on causal reasoning, or if it merely relies on brittle shortcuts that fail in rare, safety-critical situations. The paper addresses the question: Do driving agents truly learn the core prediction and planning skills required for robust driving, or do they rely on spurious correlations?

Methodology
The authors investigate the internal representations of driving models—specifically Behavior Cloning (BC), Reinforcement Learning (RL), and Imitation Learning (IL)—across varying data scales. They employ three primary analytical techniques:

1. Linear Probing: To assess what information is encoded in the model’s internal layers, the authors train linear classifiers on intermediate representations to predict future positions. They perform two types of probing:

   • Surrounding-Vehicle Probing: Decoding the future positions of nearby agents to evaluate prediction capability.
   • Ego Probing: Decoding the ego vehicle’s planned future positions to evaluate planning capability.
   • They compare these probes against baselines trained on raw inputs to measure the "probing gain" (LP – Raw), indicating how much additional semantic information the model learns beyond raw features.

2. Perturbed Closed-Loop Evaluation: To test dependency on surrounding vehicles, the authors randomly remove surrounding vehicles from simulation scenes. If a model relies on spurious correlations rather than true interaction reasoning, performance should degrade even when the driving task theoretically becomes easier (single-vehicle driving).

3. Causal Intervention: To establish a causal link between prediction and planning, the authors intervene on the internal representations of surrounding vehicles. By modifying the latent features of a specific surrounding vehicle to correct a misprediction or induce a specific trajectory, they observe whether the ego vehicle’s planned trajectory adapts accordingly. This tests the adaptiveness and recovery capabilities of the planner.

Key Contributions

• Scaling Analysis of Prediction/Planning: The paper demonstrates how prediction and planning capabilities emerge and plateau as training data scale increases, distinguishing between improved behavioral heuristics and genuine representational learning.
• Perturbed Simulation Protocol: The introduction of a surrounding-vehicle removal protocol reveals that driving models often perform worse when surrounding vehicles are removed, suggesting a spurious reliance on the presence of other agents rather than robust interaction reasoning.
• Near-Collision Analysis and Causal Intervention: The study shows that models often fail to form timely predictions of safety-critical vehicles during near-collision events. Crucially, it demonstrates via causal intervention that correcting internal prediction errors leads to safer and more appropriate ego planning trajectories.

Results

• Scaling Laws: Both BC and RL models follow power-law scaling relationships in closed-loop metrics. However, linear probing reveals that while models learn to ignore irrelevant distant vehicles as data scale increases, they still struggle to identify and prioritize truly safety-critical vehicles early enough.
• Prediction vs. Planning: BC models show improved prediction representations with scale, but these gains are concentrated on simple behaviors (straight driving) and short horizons. RL models show stronger short-horizon prediction representations but discard long-horizon information. IL models (SMART) preserve surrounding vehicle information better over longer temporal ranges.
• Spurious Dependencies: In perturbed simulations where surrounding vehicles were removed, both BC and RL models exhibited increased off-road rates and worse goal progress, despite the absence of interaction constraints. This indicates that the policies partially rely on the presence of other vehicles in a spurious way, likely due to dataset bias toward multi-vehicle scenarios.
• Late Prediction in Critical Events: Near-collision analysis reveals that models typically only increase attention to the colliding vehicle in the final 10 steps before a crash. This late recognition is insufficient for effective avoidance, leading to planning failures.
• Causal Impact of Prediction: Intervention experiments show that when incorrect surrounding-vehicle predictions are corrected in the latent space, the ego vehicle’s planning adapts. For example, correcting a mispredicted path of a surrounding vehicle steers the ego plan toward a safer, ground-truth-aligned trajectory. This confirms that internal predictions causally influence ego planning.

Significance and Claims
The paper claims that better closed-loop performance does not necessarily equate to better internal reasoning. It highlights a critical limitation: current driving policies often fail to anticipate safety-critical interactions early enough. The significance of the work lies in demonstrating that internal prediction errors are directly linked to planning failures and that correcting these errors can improve safety. The authors suggest that future improvements in autonomous driving require not just larger datasets, but methods that explicitly encourage models to learn precise, timely, and causal representations of surrounding vehicle dynamics, rather than relying on nominal heuristics.