LAD and RAD: The Future of Real-Time Autonomous Planning?

The world of autonomous vehicles just got a shake-up with the introduction of LAD and RAD, two new planning architectures designed to enhance real-time decision-making. LAD, a language-action planner, claims to produce a motion plan in a single forward pass at a rate of about 20 Hz. It can even generate textual reasoning alongside a motion plan at around 10 Hz. This marks a significant leap in speed, achieving roughly three times lower latency compared to previous driving language models.

LAD: A New Frontier in Real-Time Planning

LAD's ability to operate in real-time closed-loop deployments is a major shift. By setting a new learning-based state of the art on benchmarks like nuPlan Test14-Hard and InterPlan, LAD demonstrates that fast, efficient planning is no longer a pipe dream. But this isn't just about speed. The real magic lies in its architecture, which is designed to be interruptible. That means dynamic environments won't easily trip it up, a essential factor for real-world applications.

Enter RAD: The Rule-Based Challenger

While LAD takes a learning-based approach, RAD offers a rule-based counterpoint. RAD is engineered to tackle the structural limitations found in previous models like PDM-Closed. Its performance on the same benchmarks shows that rules can still hold their own. In fact, RAD achieves state-of-the-art results among rule-based planners. But why stop at rules or learning when you can have both?

Hybrid Planning: The Best of Both Worlds?

The combination of RAD and LAD into a hybrid system is where things get interesting. By fusing rule-based reliability with the adaptability and explainability of language, this hybrid model captures the strengths of both approaches. It's a bold move that addresses the often binary debate in the industry: should we stick with rules, or is learning the way forward? The hybrid system suggests we don't have to choose.

If the AI can hold a wallet, who writes the risk model? In the case of autonomous driving, it's not just about who trains the model. It's about which model can adapt and explain its decisions in real time. The intersection is real. Ninety percent of the projects aren't. But LAD and RAD might just be part of that essential ten percent that defines the future of autonomous planning.

Why Should We Care?

So why does this matter? Autonomous vehicles aren't just a tech curiosity. They're poised to redefine transportation, logistics, and even urban planning. If LAD and RAD can deliver on their promise, they'll be important players in this transformation. But let's not get ahead of ourselves. Show me the inference costs. Then we'll talk.