A driving model learns to follow requested speeds
Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving · arXiv:2610.36079
In plain words
Driving demonstrations show a model one way to travel through each scene, even when several speeds might work. Their speed choices can also be inconsistent.
The paper introduces CLAP-AD, a planning head that predicts a path and separately checks possible speeds for collisions and whether the car can stop at lights and stop signs. In a test across 48 routes, it helped one driving model follow requested speeds more closely than the model without the new head. The test does not show that every road rule is covered.
The research
A driving model trained from demonstrations sees one speed choice for each scene, even when several could work. If those choices are noisy or ambiguous, can the model follow a requested speed while still checking whether it is safe?
CLAP-AD separates path prediction from speed scoring, letting a driving model select a requested speed from options that pass its safety checks.
Can a driving model handle several valid speeds when its demonstrations show only one choice per scene? CLAP-AD separates path prediction from speed selection, scoring candidate speeds for collisions and stopping at traffic lights and stop signs.
In a closed-loop test across 48 routes, TFv6 with CLAP-AD followed requested speeds more closely than TFv6. The baseline could not accept requests, so its score was only a reference for how often its fixed speed matched them. The result supports test-time speed control in this setting.
These results come from three specific evaluation settings, and the scores are not directly comparable across them. The scoring covers collisions, traffic lights and stop signs, while the candidate speeds focus on urban driving; it does not test every road rule, maneuver or real-world condition.
What the evidence shows
- One scene, many speeds
Closed-Loop-Aware Planning for Autonomous Driving (CLAP-AD) is a plug-in planning head for end-to-end driving models. It addresses a mismatch in imitation learning: demonstrations show one behavior for a scene, although several speeds may be valid.
If those speed choices are noisy or ambiguous, they can also give the model inconsistent supervision.
- Score speeds separately
CLAP-AD learns the path by imitating demonstrations but scores candidate speeds separately for collision risk, red-light stopping and stop-sign stopping. Its scores come from targets derived from logged driving data and rules for these criteria. At test-time, a requested speed can guide selection among speeds that pass all the criteria.
- Requests are followed more closely
The Bench2Drive-Speed test measures how closely a vehicle’s actual speed follows a requested speed across 48 routes. TFv6 with CLAP-AD scored 82.53 Speed-Adherence points, compared with 69.64 for TFv6.
The baseline cannot receive a speed request, so its score is only a reference for how often its fixed speed happened to match.
- Gains vary by backbone
On 220 closed-loop Bench2Drive routes, CLAP-AD raised UniAD’s Driving Score from 45.81 to 74.29 and VAD’s from 42.35 to 75.04. Both were trained with Think2Drive demonstrations.
TFv6, trained with rule-based PDM-Lite demonstrations, instead fell from 94.70 to 92.30, an exception to the gains on the other two backbones.
- The head matters too
A 10-route Dev10 ablation found that multi-criteria scoring with UniAD’s original planning head scored 40.50 Driving Score points, while using CLAP-AD’s head with the same scoring objective scored 59.44. This suggests the head matters to how the scoring objective works in that setup.
On NAVSIM v2 navhard, CLAP-AD also improved LTFv6’s EPDMS from 31.9 to 38.0, with scoring restricted to collision risk.
Deep
- Original paper authors
- Emad Aghajanzadeh, Thomas Schwarzfischer, Marc Kaufeld, Moritz Thoma, Ryan Pégoud
- Submitted to arXiv
- License
- arXiv non-exclusive