Tsinghua's CLEAR Framework Hits 93.7 PDMS Running Self-Driving AI at 99 FPS
CLEAR replaces diffusion's slow denoising with a single-step VAE drift guided by a tiny LLM, hitting 93.7 PDMS on NAVSIM v1 at up to 99 FPS
Takeaways
- New SOTA on NAVSIM v1: CLEAR achieves 93.7 PDMS, surpassing Drive-JEPA (93.3) and DriveSuprim (93.5) while running at up to 99 FPS.
- Single-step generation: Replaces diffusion's multi-step denoising with one forward pass through a VAE latent space, eliminating the core latency bottleneck.
- LLM as feature extractor: Fine-tuned Qwen 3.5 0.8B is used only for its hidden states -- never its text output -- to guide trajectory diversity and selection.
- Adaptive compute allocation: An LLM-driven scheduler picks how many candidates to generate and how diverse they should be, spending more compute on complex intersections and less on highway cruising.
- No LiDAR or 3D annotations needed: Uses only a single front-view camera and Drive-JEPA visual features, making it practical for low-cost deployment.
- Weaknesses remain: Lags on lane keeping and traffic light compliance in NAVSIM v2; discrete scheduling grid may miss optimal configurations; no code released yet.
End-to-end autonomous driving has a dirty secret: the generative models that produce the most realistic, multi-modal driving plans are also far too slow to run in a real car. Diffusion models can capture the full spectrum of how a human might navigate a crowded intersection, but their iterative denoising process can require tens to hundreds of forward passes per decision. The high latency is due to repeated inference steps required for denoising -- 2D path planning using diffusion operates at just 2.5 Hz, while autonomous vehicles typically require path planning at 20 Hz or faster. That gap is a safety problem, not just an engineering inconvenience.
A new paper from researchers at Tsinghua University and collaborators proposes CLEAR (Cognition and Latent Evaluation for Adaptive Routing), a framework that keeps the diversity of generative planning while sidestepping the latency wall entirely. The result is a state-of-the-art score of 93.7 PDMS on the NAVSIM v1 benchmark, running at up to 99 FPS -- without dense 3D geometric annotations or iterative sampling.
The core tension in autonomous driving planners
Urban driving is inherently multi-modal. At an unsignalized intersection, yielding, proceeding, and merging are all valid maneuvers simultaneously. End-to-end autonomous driving has traditionally relied on regression-based approaches that predict a single trajectory to mimic expert behavior. The problem: averaging across incompatible modes produces physically implausible paths that straddle two valid maneuvers at once.
While diffusion models successfully capture diverse driving behaviors, their iterative denoising process incurs unacceptable latency for safety-critical deployment. Accelerated samplers like consistency models and flow matching help, but they sacrifice trajectory quality in high-dimensional control spaces. The field has been stuck choosing between diversity and speed.
CLEAR's insight is that most denoising steps in a diffusion chain are doing something mundane: removing Gaussian noise. The semantically meaningful shift -- from an unconditional distribution to a scene-conditioned one -- can theoretically happen in a single step. The team grounds this in the theory of