DiffusionDrive

What It Is

DiffusionDrive is an end-to-end autonomous driving model that uses a truncated diffusion policy as its planner head.
It was proposed by Liao et al. and accepted to CVPR 2025 as a Highlight.
The method addresses the main practical limitation of diffusion-based driving policies: their denoising chains are usually too long to meet real-time inference budgets in open-world traffic.
DiffusionDrive demonstrates that diffusion planning can be both multi-modal and fast enough for driving when the diffusion schedule is truncated and seeded from learned anchors.

Treat ego planning as a multi-mode action distribution and learn it with a diffusion policy.
Replace pure Gaussian noise initialization with prior multi-mode anchors that already cover the typical maneuver set, then run a short denoising schedule from those anchors instead of from scratch.
Truncate the diffusion schedule so the policy only needs a handful of denoising steps, often as few as two, while still producing diverse and high-quality trajectories.
Use a cascade diffusion decoder so each denoising step interacts efficiently with scene context features rather than treating the score model as an isolated network.

Input: multi-view surround camera images with intrinsics, extrinsics, ego pose, timestamps, and optional navigation command.
Conditional input to the policy: scene context features from the upstream perception stack.
Training input: ego-trajectory supervision and the standard NAVSIM planning-oriented evaluation labels.
Output: a set of candidate ego trajectories sampled from the truncated diffusion policy, plus a selected trajectory.

Perception backbone extracts multi-view image features; the paper uses a ResNet-34 backbone for the aligned comparison setting.
A scene context encoder turns image features into a compact representation consumed by the policy decoder.
Multi-mode anchors are learned or chosen so they cover dominant ego-maneuver clusters, then perturbed into the truncated diffusion starting distribution.
A cascade diffusion decoder runs a small number of denoising steps, each conditioned on scene context features.
The decoder emits multiple trajectory samples per inference call; selection logic chooses the final plan.

Main benchmark: NAVSIM planning-oriented evaluation built around the OpenScene data and the PDM-style scoring metric.
Reported result: 88.1 PDMS on NAVSIM with an aligned ResNet-34 backbone, setting a new record at submission time without specialized backbones or tuning tricks.
Inference speed: 45 FPS on a single NVIDIA RTX 4090 with two denoising steps, which is roughly a ten-times reduction in denoising steps compared to a vanilla diffusion policy at similar quality.
Qualitative results on challenging scenarios show diverse plausible trajectories rather than mode collapse.
Comparison against vanilla diffusion policy and against regression and classification-based planners isolates the contribution of truncation, anchors, and cascade decoding.

Truncated diffusion makes diffusion-based driving policies practical for real-time inference on production-class GPUs.
Multi-mode planning naturally captures maneuvers such as nudge-left, brake-and-yield, or lane-change without collapsing to a single regression mean.
Cascade decoder design uses scene context per step, which improves grounding compared to score networks that ignore conditioning detail.
Strong NAVSIM result with a small ResNet-34 backbone shows the gains come from the policy design, not from a heavier encoder.
Provides a clean baseline for combining diffusion planners with sparse or VLM-style perception stacks.

NAVSIM is open-loop and does not capture all closed-loop safety failure modes; reported numbers should not be read as full validation.
Multi-mode outputs require a selection step; weak selection logic can pick a low-probability but visually plausible mode.
Truncated schedules make the policy more dependent on the quality of the anchor distribution; if anchors miss a needed maneuver class, the policy may not recover with so few denoising steps.
Diffusion-based planners remain harder to certify than constrained optimization planners because their decision boundaries are implicit.
Real-time numbers are reported on a desktop-class RTX 4090; in-vehicle compute may require additional optimization or larger denoising budgets.

Road AV: primary intended domain; NAVSIM evaluation targets urban driving with realistic agent interactions.
Airside: relevant only as a pattern; airside trajectory generation must respect aircraft separation, wing clearance, and stand geometry, which are not represented in NAVSIM. Any airside use needs a safety filter on top of the diffusion samples.
Warehouse, yard, port, mining, construction, agriculture, delivery robot, campus: usable where multi-mode maneuver choice helps, such as around humans and equipment, but always paired with a deterministic safety envelope or rule layer.
General: best treated as a planner pattern that complements a separate safety case rather than as a stand-alone policy.

Reproduce the reported NAVSIM result with the aligned ResNet-34 backbone before changing perception or backbone.
Choose and audit the multi-mode anchors per deployment domain; anchors learned from public driving data may not cover domain-specific maneuvers such as taxiing or yard turns.
Measure inference latency on the target compute with the production perception stack; do not rely on RTX 4090 numbers.
Pair the diffusion policy with a safety filter, runtime monitor, or rule-based planner that can reject unsafe samples regardless of policy confidence.
Log sampled trajectories, selected trajectory, and selection scores per frame for offline auditing and disengagement analysis.
Evaluate in closed-loop simulation and in shadow mode on a fleet before drawing planning conclusions from open-loop scores alone.