Cars that anticipate a lane change or turn a few seconds before it happens give a planner time to respond smoothly instead of braking hard. Today's anticipation models map driver head motion, gaze and traffic context straight to a probability such as "left lane change, 0.82." Their documented failures are cases where that direct mapping is ambiguous. A driver drifts toward the lane line and then aborts because the gap closes. A head turn is a blind-spot check in one case and a glance at the radio in another. The model also cannot say why it expects the maneuver. We propose that the anticipator first write out a short, structured chain of intermediate judgments and only then predict the maneuver class and time-to-maneuver from that chain. The chain answers four questions: does the driver show desire (a mirror check, speed pressure from a slow leader), is the target lane feasible (gap acceptance, time-to-collision), what is the motive, and has the driver committed? This is chain-of-thought prompting, meaning a model generating intermediate reasoning steps before its answer, shrunk to about ten discrete tokens. Every token is labeled automatically from geometry and annotations, so no language model writes the rationale.
The claim is not "chain-of-thought helps driving." Ego-planning work has already tested that. The claim is narrower and falsifiable. When the final decision must attend to its own decoded prerequisites, three things should improve over a model trained on the same intermediate labels as parallel side outputs. False alarms on aborted maneuvers should drop. Distraction glances should be separated from intent checks. And the prediction should become intervenable: if you overwrite "gap feasible" with "gap closed," the predicted lane change should retract. If the sequential-conditioning version does no better than the parallel one, the honest conclusion is that the benefit came from extra supervision, not from reasoning. The experiments are built to show which.
What exists and where this sits
The anticipation literature supplies the signals and the targets. Car that Knows Before You Do fuses head-motion histograms with road context through a hidden intent state in a hidden Markov model. Brain4Cars replaces this with separate inside and outside recurrent networks and an exponential loss that penalizes errors more as the maneuver nears. It also names the distinction between intentional head turns and distraction as unsolved. Early Anticipation of Driving Maneuvers adds multi-view video and eye-tracked gaze on the Driving Action Anticipation Dataset (DAAD), and flags "cause" labels as missing. All three motivate the proposal: they show driver cues lead vehicle motion by seconds, and they leave ambiguity and explanation open.
Three papers supply the feasibility side, and we reuse their quantities as chain fields rather than as raw inputs. Dual-View Interaction-Aware Lane Change Prediction scores destination-lane safety from time-to-collision variants and a spatial risk index. Artificial Potential Fields detects intent from the potential-energy difference between lanes. Shou et al., summarized in the file, show that longitudinal gaps and speed differences alone predict intent well ahead of onset. Attention-based Lane Change Prediction documents aborted "fake" maneuvers as a false-alarm source. Beyond the Data Imbalance is a constraint on the design: naive models fall into a straight-line bias with zero turn recall, so any chain must be trained with rebalancing.
The closest connecting work applies chain-of-thought to driving, but always to a different question. DriveCoT predicts discrete reasoning fields before the ego car's speed and waypoints in simulation. LCDrive replaces text with latent action and world-model tokens for ego trajectories. FastDriveCoT decodes independent template fields in parallel as a dependency graph to meet 10 Hz deadlines; we borrow this decoding trick. CoT-Drive is nearest. It distills GPT-4 scene descriptions into a small language model for forecasting surrounding-vehicle trajectories. However, it forecasts coordinates from motion already underway, uses no driver cues, and admits it struggles where history contradicts future intent. That last case is exactly the pre-maneuver window. Accident anticipation with LLMs turns predictions into verbal alerts after the fact, and its reasoning does not feed the decision.
Our design differs from all of these in two ways. Its rationale fields are computed quantities, not language-model prose, which addresses the geometric-illiteracy and post-hoc-rationalization critiques in the action-grounded reasoning survey. And we test whether the chain causally controls the output. The template idea comes from Reflect, Not Reflex, whose prerequisite and consequence inference step improved dialogue. That paper is also a constraint: small models gained only when supervised on intermediate inferences, not when prompted. This is why we supervise. The intervention test borrows its logic from A Case Study of How Intuition and Reasoning Interact, where prefilling a reasoning chain reversed the model's decisions. We apply the same test to ask whether the final prediction follows the chain.
Core Claim
Let $x_t$ be the sensor history up to time $t$. Let $c_t$ be the decoded prerequisite chain: desire evidence, feasibility, motive and commitment, each a short discrete token. Let $(m, \tau)$ be the maneuver class and time-to-maneuver. The model factorizes
$$p(m,\tau \mid x_t) = \sum_{c} p(c \mid x_t)\, p(m,\tau \mid c, x_t),$$
and in practice decodes the most likely chain and then conditions on it. Desire and feasibility depend on disjoint evidence: in-cabin cues versus exterior traffic. They can therefore be decoded in parallel as in FastDriveCoT, and motive and commitment follow them. The hypothesis is that conditioning on $c$ matters most when desire and feasibility disagree. A driver who looks but has no gap should produce "desire yes, feasible no, commit no." A drift without a mirror check toward a closing gap should not trigger an alert.
Research Program
We first build the chain labels. On HighD (drone-recorded highway trajectories), feasibility is labeled from target-lane time-to-collision and gap length at each step. Motive is labeled from a slow leader, a lane drop or a merge. We mine aborted lane changes as tracks that drift laterally toward a boundary and return without crossing; HighD's 25 Hz sampling makes this mining reliable. HighD has no in-cabin view, so desire there comes only from longitudinal pressure.
The Brain4Cars dataset (in-cabin and road video, 700 annotated events) adds driver cues. Desire is labeled from head-yaw events toward the mirror or target side, derived from the facial landmarks. Feasibility comes from the outside lane and vehicle annotations. We also mine "looked but did not go" segments from straight-driving events with large head turns, which gives a natural test set for distraction versus intent. DAAD adds gaze-based desire labels and its rural and unmarked-road cases.
We train a small temporal encoder with a chain decoder of about ten tokens and a decision head, using Brain4Cars's exponential loss and class rebalancing. We measure latency against a 100 ms budget.
Baselines
The do-nothing reference is the model's own class prior plus a kinematic trigger that fires at lateral-velocity onset. It defines zero anticipation lead time, and any learned method must beat it in seconds of warning.
The strongest existing method on driver-cue data is the Brain4Cars sensory-fusion recurrent network, which reports 90.5% precision, 87.4% recall and 3.5 seconds of anticipation on the Brain4Cars dataset. The AIO-HMM reference on that dataset reports 77.4% precision and 71.2% recall. On HighD, the strongest number is the driving-preference-aware model's macro F1 of 0.9733 from A Driving-Preference-Aware Framework. On DAAD, M²MVT reports 65.59% accuracy and 66.72% F1. We found no reported result on Lacuna for false-alarm rate on aborted lane changes, or for separating distraction from intent, for any method. We will establish those baseline numbers ourselves.
For fairness, every baseline is retrained on our splits with the same rebalancing. Each gets an equal hyperparameter-search budget measured in GPU-hours, and each is given the computed time-to-collision, gap and potential-field features as extra inputs. A feature-augmented Brain4Cars network is therefore the true strongest competitor. Beating it means one of two things: higher precision at equal or greater lead time, or a significantly lower abort false-alarm rate at matched recall. Gains on the headline metrics alone are not the target.
The key ablation is a multi-task model with the identical encoder and identical chain labels, where the chain fields are predicted by parallel heads that the decision head cannot see. A second ablation feeds the decision head the ground-truth chain, which gives an upper bound on how much a perfect chain could help. A third shuffles chain tokens at test time to check that the decision actually depends on them.
We add a CoT-Drive-style variant. An open vision-language model writes free-text rationales, which are distilled into the same small model. This tests computed rationales against language-model rationales.
Why not just…?
Why not just add auxiliary heads for gap, mirror check and motive (a concept bottleneck)? This is the real threat. Multi-task supervision alone may deliver most of the gain. The parallel-heads ablation exists to settle it. If chain-conditioned decoding does not beat it on aborted maneuvers and look-but-no-go segments, we will report that the benefit is supervision. Even in that case, intervenability separates the two designs. An auxiliary-head model offers no handle to edit "gap closed" and watch the prediction respond.
Why not just feed the safety features into the Brain4Cars network? Dual-View shows such features help forecasting, so this baseline will be strong. That is why it is our main competitor. The chain's argued advantage is the explicit interaction term "desire present but infeasible." A recurrent network must discover that conjunction from few aborted examples. The abort and distraction test sets measure exactly this.
Why not just prompt a large vision-language model with chain-of-thought? Large models write fluent rationales, but the survey and LCDrive document weak metric grounding and latency problems. Reflect, Not Reflex shows that small deployable models do not benefit from prompting alone. We still run a zero-shot large-model reference on a subset, so readers can see its accuracy and latency.
Proposed Experiments
We report precision, recall, F1 and mean time-to-maneuver on Brain4Cars and DAAD, and macro F1 and lead time on HighD. Beyond those, three checks decide the claim.
The first is the false-alarm rate on mined aborted lane changes and on look-but-no-go segments, compared at matched recall against the feature-augmented Brain4Cars network and the parallel-heads ablation.
The second is an intervention test. We overwrite one chain field at test time, for example setting feasibility to infeasible, and measure how often the prediction flips in the direction the edit implies. A flip rate near chance would mean the chain is decorative, as the post-hoc-rationalization critique predicts. We also evaluate motive accuracy against held-out manual labels on a few hundred DAAD and Brain4Cars events, which gives the "cause" output that the DAAD authors ask for.
The third is latency on a single GPU with parallel decoding of the desire and feasibility fields, which must fit inside 100 ms. We will judge the proposal successful only if the chain beats parallel heads on aborts, flips under intervention, and meets the latency budget, all without losing lead time.
References
- Car that Knows Before You Do: Anticipating Maneuvers via Learning Temporal Driving Models
- Brain4Cars: Car That Knows Before You Do via Sensory-Fusion Deep Learning Architecture.
- Early Anticipation of Driving Maneuvers
- Dual-View Interaction-Aware Lane Change Prediction for Autonomous Driving
- Automatic Detection Method of Lane-Changing Intentions Based on Relationship with Adjacent Vehicles Using Artificial Potential Fields
- Attention-based Lane Change Prediction
- Beyond the Data Imbalance: Employing the Heterogeneous Datasets for Vehicle Maneuver Prediction
- DriveCoT: Integrating Chain-of-Thought Reasoning with End-to-End Driving
- Latent Chain-of-Thought World Modeling for End-to-End Driving
- Accelerating Structured Chain-of-Thought in Autonomous Vehicles
- CoT-Drive: Efficient Motion Forecasting for Autonomous Driving with LLMs and Chain-of-Thought Prompting
- When, Where, and What? A Benchmark for Accident Anticipation and Localization with Large Language Models
- Beyond Textual Chain-of-Thought: A Survey on Action-Grounded Reasoning in Autonomous Driving
- Reflect, Not Reflex: Inference-Based Common Ground Improves Dialogue Response Quality
- A Case Study of How Intuition and Reasoning Interact in Language Models
- A Driving-Preference-Aware Framework for Vehicle Lane Change Prediction