PaperlayerView exact revision on arXiv

Source-linked AI summary

Gating Before Commitment: Anticipating Intent Divergence to Prevent Post-Interaction Decision Failures in Autonomous Driving

Cong Xu, Ravi Sankar

arXiv:2608.26074v1cs.ROcs.AI

TL;DR

Intent misinterpretation during vehicle interactions can turn transient perception disturbances into planning failures. The paper evaluates a language-guided decision layer that computes smoothed intent-geometry divergence and gates maneuvers before commitment. Across a real off-road departure and four crash clips, gating repaired the plan, while abstention semantics reduced false triggers and preserved intent-level detections.

  • Problem

    Intent misinterpretation during vehicle interactions causes recurring planning failures, including failures attributed solely to perception limits.

  • Method

    A language-guided intent module reads structured descriptors, computes smoothed intent-geometry divergence, and gates and repairs planned maneuvers upstream of a corridor envelope.

  • Results

    Gating was the only layer that repaired the plan; abstention semantics reduced false triggers from 1.54 to 0.341 per minute while preserving every intent-level detection.

  • Takeaways & Limitations

    The evidence supports the gating mechanism, with the model contributing fastest detection on these failures and an uncertainty veto on the geometric rule.

  • Takeaways & Limitations

    The implementation is a surrogate stack, calibration does not transfer to the out-of-domain failure clips, and five-second crash clips cannot assess recoverability.

Abstract

from arXiv · show

Intent misinterpretation during vehicle interactions causes recurring planning failures. We study a decision layer in which a language-guided intent module reads structured descriptors, computes a smoothed intent-geometry divergence score, and gates the planned maneuver before commitment, upstream of a corridor envelope. On a replayed off-road departure and four crash clips under a frozen, disclosed implementation, gating is the only layer that repairs the plan: on the main case it fires 72 ms after the drift onset but 161 ms before the corridor exit, keeping the trajectory in the corridor in all ten replays. The first calibration draws nine false triggers in 5.9 minutes, each from scoring uncertainty as half a conflict; a preregistered redesign treating uncertainty as abstention cuts this to 0.341 per minute. Two ablations bound the model's contribution: the full score detects fastest on four of five failures under the deployed eligibility, three of five against the unvetoed rule (000871 by one cycle; 000228 by a pre-onset fire on an uncertain stretch that five clips cannot classify as signal or coincidence; dropping the confidence term costs two detections), while on in-domain tracks at equal false positives the geometric rule more than triples its detection. The evidence supports the gating mechanism; the model's demonstrated roles are the fastest detection on these failures and an uncertainty veto on the geometric rule.

I. INTRODUCTION

Interactive driving failures can arise when a planner misreads intent during post-interaction phases, turning transient perception disturbance into an unrecoverable maneuver. The paper proposes anticipating intent-geometry divergence and gating the plan before commitment, then evaluates the mechanism and its uncertainty redesign.

  • Interaction phases such as passing, merging, and lane changes remain a primary source of catastrophic decision failures.
  • 167 ms after the oncoming vehicle clears the ego plane, measurable drift begins, and the vehicle leaves the road corridor within 400 ms.
  • The decision layer infers intent hypotheses, computes intent-geometry divergence, gates and repairs maneuvers, and uses an intent-conditioned safety envelope as a reactive backstop.
  • Intent-level failures reframe interactive-driving errors as misinterpretations of ego and other-vehicle intent, rather than solely perception failures.
  • 1.54 to 0.341 false triggers per minute followed an abstention-semantics redesign with dual-threshold hysteresis while preserving every intent-level detection.
  • The full score detects fastest on four of five failures under deployed eligibility and three of five against the unvetoed rule, while geometry-only detection more than triples in-domain at equal false positives.

II. RELATED WORK

Prior driving systems largely predict trajectories or react after unsafe states emerge, whereas this paper places a language-guided intent module upstream to issue plan-correction signals. Its text-only interface operates on structured scene descriptors rather than richer visual representations.

  • Trajectory forecasting, game-theoretic planning, and interaction-aware predictors inform planners but generally output point predictions rather than uncertain intent hypotheses.
  • Scenario analysis, formal verification, minimum-risk fallbacks, and online plan repair address aftermath or observed deviations rather than anticipating intent divergence.
  • The language model functions as an intent anticipation module upstream of planning and issues correction signals rather than control commands.
  • The decision loop uses a structured, text-only interface derived from a deterministic semantic serialization of a frozen perception stack.
  • Unlike richer monocular reconstruction and layered language-vision maps, the measured implementation uses a lean detector, tracker, lane estimator, and pinhole ground-plane model.

III. FRAMEWORK

The framework converts monocular video into structured interaction descriptors, predicts categorical intent with calibrated confidence, and compares semantic intent with the planner’s geometric motion mode. That divergence signal supports downstream maneuver gating.

  • Monocular video is processed with object detection and tracking, color-based lane fits, and a pinhole ground-plane model that estimates metric range and lateral offset.
  • A lightweight digital twin renders a bird’s-eye-view map with a semi-automatically annotated road corridor.
  • Structured descriptors include relative positions, closing speed, heading offset, lane assignment, occlusion, and interaction phase.
  • 1) Intent distribution and uncertainty:: The intent module maps descriptor vector x_t to a categorical hypothesis over yield, block, merge, pass, and uncertain.
  • 1) Intent distribution and uncertainty:: The MAP label is the highest-probability intent, while calibrated confidence c_t is the only distributional quantity consumed by the gate.
  • 1) Intent distribution and uncertainty:: A fixed lookup table compares the geometric predictor’s motion mode with the inferred semantic mode to produce divergence d_t in [0, 1].

2) Intent-geometry divergence:

The decision layer smooths intent–geometry divergence and uses it to gate maneuver commitment. Version 2 treats uncertain intent as abstention and restricts Red gating to semantic hypotheses, with hysteresis added in the primary configuration.

  • The smoothed divergence score combines disagreement between predicted intent and geometry with model confidence, then feeds a conflict flag.The score is exponentially smoothed before thresholding.
  • Uncertainty as abstention: v1 assigned d(πpred, uncertain) = 0.5, treating every non-ambiguous uncertain forecast as half a conflict.This scoring choice later accounted for the nominal false triggers.
  • Uncertainty as abstention: v2 sets d(πpred, uncertain) = 0, so uncertainty contributes no disagreement while preserving the confidence term for semantic hypotheses.The ambiguous entry was already zero in v1.
  • Uncertainty as abstention: An uncertain cycle can raise at most Yellow; Red, gating, and repair require that cycle’s MAP label to be a semantic hypothesis.Uncertain cycles can still carry residual disagreement memory into a later semantic cycle.
  • Hysteresis: The primary v2b configuration adds dual-threshold hysteresis, clearing the flag only when the smoothed score falls below θoff = θon − 0.10.v2a uses abstention alone, while v2b adds hysteresis.

C. Intent-Aware Decision Correction and Plan Gating

The correction layer intervenes before execution by restricting admissible maneuvers and repairing the trajectory, while an intent-conditioned safety envelope remains a reactive containment backstop. The implementation uses structured monocular descriptors and a disclosed calibrated intent model under a 100 ms planning cycle.

  • Decision correction: When gt = 1, maneuver gating suppresses lateral corrections and plan repair returns the trajectory toward the lane center at a bounded lateral rate.Both interventions occur upstream of execution and commitment.
  • Safety backstop: The safety envelope rejects committed waypoints outside the corridor and tightens under Yellow or Red, but does not modify the plan in this surrogate.Its 10/10 exits measure when corridor checks would notice excursions, not controller intervention.
  • Implementation: The intent model maps fixed-schema text descriptors to labels using Qwen2.5-0.5B with a LoRA adapter and five label-token logits.Inference uses bfloat16 on a single RTX 5090 laptop GPU.
  • Training: Training uses 55,433 interaction windows labeled by deterministic future-kinematics rules without human annotation.The labels include merge, pass, yield, block, and uncertain.
  • Calibration: v2 selected θon = 0.35 with detection 0.632 and false positive 0.345, while τ, λ, and α carried over unchanged.Each configuration was frozen before experiments.
  • Runtime: The intent forward pass takes 32.7 ms median and the intent-plus-gating-plus-repair portion averages about 38 ms per 100 ms planning cycle.Perception time is excluded from reported latency.

IV. REPLAY STUDY OF A REAL-WORLD FAILURE

The replay study reconstructs a real oncoming-vehicle interaction in which a surrogate planner reproduces the observed post-interaction off-road excursion. Deterministic anchors define drift onset, corridor exit, commitment, and latency across repeated replays.

  • Study setup: A monocular rural-road replay contains an oncoming pickup interaction followed by a production-planner lateral correction that crossed the centerline and ended in a rollover.The production planner was unavailable, so the failure was replayed with a surrogate.
  • Study setup: The surrogate pure-pursuit planner follows perceived lateral state and reproduces the observed off-road excursion without correction.This configuration is the Baseline.
  • Anchors: The drift onset is defined as the first frame with lane-normalized ego offset above 1.0 m that continues growing for five valid frames.It occurs at frame 983, 167 ms after t0.
  • Anchors: The Baseline corridor exit is the first cycle when executed lateral state leaves the corridor, occurring 400 ms after t0.This is cycle 330 at t = 33.0 s.
  • Measurement: Latency is measured from t0 to the first gate or envelope event, with commitment handed to the tracker at each 100 ms cycle.Ten replays per configuration are deterministic in decision cycle, with variability reflecting computation-time jitter.

C. Configurations and Results

The main-case comparison evaluates Baseline, Intent Only, Envelope Only, and Full configurations under v1 and preregistered v2 settings. Gating is the layer that repairs the trajectory, firing before corridor exit, whereas the envelope detects the excursion only after commitment.

  • Outcome: The Baseline exits the corridor in all replays, while intent gating repairs the plan in all replays under v1, v2a, and v2b.The result is unchanged across the three gating configurations.
  • Table I: 10 replays per row compare main-case outcome and latency using drift onset and corridor exit as gate anchors.The table reports variability as one standard deviation.
  • Outcome: 10/10 envelope replays reject the committed trajectory, but the envelope acts after commitment and does not modify the plan.Baseline and Envelope arms are identical across configurations.
  • Replay anatomy: The main-case replay shows the envelope flagging the excursion one cycle before the intent module triggers Red, gating, and repair.The intent trigger follows a low-confidence pass against an abnormal committed lateral rate.
  • Latency: 72.2 ± 2.0 ms after drift onset and 161.1 ± 2.0 ms before corridor exit, the v1 gate fires before the corresponding commitment.Under v2b it fires 67.7 ± 3.7 ms after drift onset and 165.6 ± 3.7 ms before corridor exit.

D. Failure Mode Analysis

The main-case failure follows corrupted post-pass perception: the planner commits a geometrically abnormal lateral maneuver, while the intent gate detects and repairs it before corridor exit. The redesigned implementation keeps gating active because release requires confident, geometry-consistent cycles.

  • Failure anatomy: One cycle after commitment, a low-confidence pass hypothesis conflicts with a crosses lane forecast, causing the smoothed divergence to cross threshold and trigger gate-and-repair.The logged sequence is: commitment, abnormal lateral rate, intent scoring, then Red, gate, and repair.
  • Failure anatomy: Beyond roughly 50 m, pixel-noise-dominated ranging produces implausible closing speeds upstream of some uncertain outputs.A v1 cycle carried 299.7 m/s; such values are marked rather than treated as physical speeds.
  • Failure anatomy: The v2b trace remains above θoff through the replay, so lateral corrections stay gated until a run of confident, geometry-consistent cycles permits release.
  • Acceptance boundary: The preregistered v2 redesign failed its all-five pre-drift criterion because the main-case gate followed drift onset by 67.7 ± 3.7 ms, despite preceding corridor exit by 165.6 ± 3.7 ms.The criterion was retained as failed rather than rewritten retroactively.

B. Additional Interaction Clips

Four screened crash clips extend evaluation beyond the main case, separating intent-level failures from post-interaction boundary checks. The redesigned gate preserves or improves intent-level detection but cannot restore containment when clips begin at the corridor boundary.

  • Clip selection: Four clips passed screening: 000754 and 000871 were typed as intent-level failures, while 000676 and 000228 were post-interaction loss-of-control boundary checks.
  • Detection results: Under v2b, every intent-level failure is detected with equal or better lead, including 000754 improving from 700 to 1000 ms and 000871 reaching 1400 ms.Boundary check 000676 is retained with an 800 ms lead; 000228 flags 1.1 s after onset.
  • Recoverability: On the 5 s pre-crash cuts, the ego is already at the corridor boundary when the gate fires, so repair mitigates but cannot restore containment in either round.Containment recovery was 0/4 in both rounds, separating detection lead from recoverability.
  • Evaluation scope: The nominal set contains 703 cycles over 5.86 minutes, while the highway probe contains 4,792 cycles over 8.0 minutes and is reported separately.The nominal corpus is a disclosed proxy because zero-interaction scenes are essentially absent in urban nuScenes.
  • False-trigger results: The v1 nominal rate was 1.54 false triggers per minute, whereas the preregistered v2 redesign reached 0.341 per minute within its bound.Abstention semantics account for the reduction; hysteresis mainly stabilizes the raised flag.
  • False-trigger results: The out-of-domain highway probe produced 25.4 false triggers per minute for v1 and 7.5 for v2b, exceeding the usability bound because perception and semantic inputs were out of calibration domain.

D. What the Language Model Contributes

The ablations test whether the language-guided score adds value beyond geometry and confidence terms. The full score is fastest on these replay failures, but matched-false-positive in-domain tracks favor the geometric rule, while uncertainty vetoes suppress nominal model-free triggers.

  • Ablation design: The corrected replay varies only the scoring term while holding the EWMA, threshold, and other configuration elements fixed.The variants are full, d-only without the (1 −ct) term, and geometry-only.
  • Replay ablations: The full score detects fastest on four of five failures under deployed eligibility, but on 000871 geometry-only fires one cycle earlier.The earlier identical-cycle result came from a replay-driver defect that was later corrected.
  • Replay ablations: Against the unvetoed rule, the full score detects fastest on three of five failures; 000228’s pre-onset trigger occurs on an uncertain stretch.
  • Replay ablations: Dropping the (1 −ct) term makes d-only fail to fire on either boundary check under deployed eligibility.
  • In-domain tracks: At matched false positives of 0.093 on nuScenes tracks, the full score detects 0.234 of conflicts, while the geometric rule detects more than three times as many.The matched comparison is the relevant one because model-driven variants otherwise fire more readily.
  • In-domain tracks: At each variant’s own best point, the geometric rule has Youden J of 0.683 versus 0.287 for model-driven scoring.
  • Control: In the merge control, Yellow rose in 6/10 scenes in both rounds, while Red fell from 6/10 under v1 to 4/10 under v2b.Red selectivity improved but remained partial.

VI. LIMITATIONS

The evaluation has important scope and implementation boundaries: the latency results come from a rebuilt surrogate stack, calibration does not transfer across domains, and several sensing, labeling, and replay assumptions constrain interpretation.

  • Implementation and evaluation scope: The August 2026 rebuild means reported latency figures characterize a surrogate stack rather than a production planner.The study reused the same data for v2 and claims no new domains.
  • Domain boundary: Calibration and false-trigger measurements are in-domain, while the rural main case and CCD clips are outside the calibration domain without claimed transfer.A highway probe also shows that calibration does not transfer to out-of-domain perception.
  • Evaluation and sensing constraints: The five-second CCD clips are too short to assess recoverability, limiting conclusions about whether the system can recover after the observed failures.The road corridor is also semi-automatic, and monocular ground-plane estimation assumes camera height and field of view.
  • Model and data constraints: Weak minority-class intent labels, deterministic rule-generated labels, noisy far-range closing speeds, and monocular descriptors constrain the evidence for intent understanding.The cooperative V2V extension is not evaluated.
  • Residual scoring limitation: The remaining confidence-as-conflict pathway buys two of five detections but also accounts for both surviving nominal triggers.The v2 redesign removed the uncertain-label route but did not eliminate effects from merely unconfident semantic hypotheses.
Loading 2608.26074v1…