Two New Planners Try to Solve Autonomous Driving's Hardest Problem: Uncertainty at Speed

The second problem is uncertainty quantification. A planner that produces a single confident trajectory when the situation is genuinely ambiguous is dangerous. What you want is a system that knows what it does not know, and that communicates this uncertainty in a way that can inform downstream decisions, whether that means slowing down, requesting human intervention, or simply hedging across multiple candidate trajectories.

Most prior work addresses one of these problems well and the other poorly. LUNA-AD and ConsistencyPlanner are, in different ways, attempting to address both simultaneously.

What is LUNA-AD actually doing?

LUNA-AD's core contribution is what the authors call a tri-system architecture. The first component is a multi-agent analytical system that generates what they describe as "uncertainty-aware decision-making demonstrations" by exploring diverse hypotheses about a driving scenario. The second is a dual-head lightweight model, distilled from the first, that can run efficiently at inference time while still producing both a distribution over decisions and a natural-language explanation for those decisions. The third is a reflection-driven lifelong learning mechanism that uses closed-loop feedback to refine the system's candidate decisions and rationales over time.

The distillation step is the part I find most technically interesting. The idea of training a large, expensive model to generate rich training data, then distilling that knowledge into a smaller, faster model, is not new. It appears in various forms in the knowledge distillation literature going back to Hinton et al.'s 2015 paper, and more recently in the robot learning space in work like DAgger variants and the broader imitation learning canon. What LUNA-AD is doing, though, is applying this to the specific problem of uncertainty representation: the lightweight model is not just imitating the large model's decisions, it is learning to reproduce the large model's uncertainty estimates. That is a harder target, and it is genuinely new as a framing in this specific domain, as far as I can find.

The lifelong learning component is also worth attention. Most deployed driving systems use static models, trained once on a fixed dataset. The argument for continual refinement is compelling in theory. In practice, catastrophic forgetting, the tendency of neural networks to overwrite old knowledge when trained on new data, remains a serious unsolved problem. The authors claim their reflection mechanism preserves strategic diversity, but the paper does not provide enough detail about how this is validated over long deployment horizons. This hasn't been replicated in a real deployment setting yet, and that matters.

On the nuPlan benchmark, LUNA-AD reportedly achieves state-of-the-art success rates in both reactive and non-reactive modes while reducing inference latency compared to existing knowledge-driven frameworks. NuPlan is a reasonable benchmark, though it is worth noting that benchmark performance and real-world performance in autonomous driving have a complicated relationship. The community has learned this lesson repeatedly.

What is ConsistencyPlanner doing differently?

ConsistencyPlanner takes a different technical bet. Rather than working with language models at all, it builds on consistency models, a class of generative model introduced by Song et al. in 2023 that can produce samples in a single forward pass rather than through the iterative denoising process required by standard diffusion models. This is, in principle, exactly the right inductive bias for real-time planning: you want multimodal trajectory generation (to capture the genuine diversity of plausible futures) without paying the computational cost of running dozens of denoising steps.

The paper's two main contributions are what they call Efficient Multimodal Sampling, using the fast-sampling property of consistency models to generate diverse trajectory candidates in real time, and Heterogeneous Feature Fusion, an attention-enhanced decoder that integrates scene features and action tokens into a unified representation.

The feature fusion component is, I will be honest, somewhat incremental over existing attention-based approaches in trajectory prediction. Cross-attention between scene context and agent state is a fairly standard technique at this point, appearing in papers like Scene Transformer (Ngiam et al., 2022) and Wayformer (Nayakanti et al., 2023). The novelty claim here rests primarily on how this is integrated with the consistency model sampling process, and the details of that integration are interesting, though the paper is somewhat sparse on ablations that would help isolate the contribution of each component.

The evaluation is done in the Waymax simulator, and the results show strong performance on safety metrics, particularly in what the authors describe as challenging dynamic scenarios. Waymax is a relatively new, Google-developed simulator that has some advantages over older environments in terms of realism, but it is still a simulation. The gap between simulated and real-world performance remains one of the central open questions in this field.

Why do these two approaches matter, taken together?

Actually, the research shows something interesting when you look at these two papers side by side: the field is converging on a shared set of desiderata, even when the technical approaches diverge significantly.

Both papers agree that multimodal output is necessary. A planner that produces a single trajectory is implicitly claiming certainty it does not have. Both papers agree that real-time performance is non-negotiable. And both papers, in different ways, are trying to build systems that can represent and communicate their own limitations.

This last point is, I think, the most important. The history of autonomous driving is littered with systems that were confidently wrong. The 2016 Tesla Autopilot fatality involved a system that failed to detect a white truck against a bright sky and did not know that it did not know. More recent incidents have involved similar failure modes. The technical community has increasingly recognized that uncertainty quantification is not an academic nicety; it is a safety requirement.

What is still unclear is whether the specific uncertainty representations in either paper are calibrated in the sense that matters for safety. A system can produce a probability distribution over trajectories that looks mathematically reasonable but is systematically overconfident in exactly the scenarios where it should be most uncertain. Evaluating calibration properly requires extensive real-world testing or, at minimum, carefully constructed adversarial simulation scenarios. Neither paper provides this level of analysis, which is understandable for preprints at this stage, but it is the thing I would want to see before treating these results as more than promising early evidence.

It is also worth noting that both papers are evaluated on single benchmarks. LUNA-AD uses nuPlan; ConsistencyPlanner uses Waymax. Cross-benchmark evaluation is rare in this literature, partly because different benchmarks measure different things, but the absence makes it harder to assess generalization.

What would I want to see next?

A few things, in rough order of importance.

First, calibration analysis. Both papers would benefit from explicit evaluation of whether their uncertainty estimates are well-calibrated, not just whether they improve planning performance. The Expected Calibration Error (ECE) metric, borrowed from the classification literature, has reasonable analogues for trajectory prediction, and there is growing work on calibration in autonomous driving specifically (see, for instance, Lakshminarayanan et al.'s work on deep ensembles as a baseline).

Second, for LUNA-AD specifically: the lifelong learning component needs more rigorous evaluation over longer time horizons. The sample size in the current evaluation is small relative to what you would need to make strong claims about continual learning without catastrophic forgetting. I know I am being picky here, but this is actually the most novel and potentially the most impactful part of the paper, and it deserves more scrutiny, not less.

Third, cross-benchmark evaluation. Running ConsistencyPlanner on nuPlan, or LUNA-AD on Waymax, would significantly strengthen the generalizability claims.

Fourth, and this is perhaps the most practically important: neither paper addresses the question of how these systems behave at the edge of their training distribution. Out-of-distribution robustness is arguably the central unsolved problem in learned autonomous driving, and it is too early to say whether uncertainty-aware architectures of this kind genuinely help with OOD scenarios or simply produce confident-sounding uncertainty estimates that are themselves unreliable in novel situations.

Both papers represent solid, careful work. LUNA-AD's distillation-based uncertainty propagation is a genuinely interesting technical contribution. ConsistencyPlanner's application of fast-sampling consistency models to trajectory planning is a natural and well-motivated idea that the field should explore further. Neither paper is going to change how autonomous vehicles are deployed next year. But they are the kind of incremental-but-meaningful research that, accumulated over time, tends to actually move things forward.

Mark Kowalski · Yesterday · 6 min