The Real Problem With Robot AI Isn't Intelligence, It's Movement

Does it actually work in practice?

The LMP framework achieved 65% task success across 31 real world manipulation tasks. The best baseline hit 35%.

Now, 65% is... not great? You might be wondering why I'm writing about something with a 35% failure rate. But context matters here. These are zero-shot tasks, meaning the system has never seen them before. No fine-tuning, no task-specific training. The robot just reads a natural language instruction and tries to do it.

For that setting, nearly doubling the success rate of existing approaches is genuinely significant. Though I should note, these were all tabletop manipulation problems. We don't know yet how this generalizes to more complex scenarios.

What about safety? (The other paper has thoughts.)

The second paper tackles a related but distinct problem. LAD-VF isn't about making robots move better, it's about making sure they don't do dangerous things when they move.

This is where I get a bit more skeptical, tbh. The paper's premise is that LLM-driven planning needs formal verification to catch safety violations. Fine. But their solution is to use formal verification feedback to automatically refine prompts, rather than fine-tuning the model itself.

The results are impressive on paper: success rates jumping from 60% to over 90% in robot navigation and manipulation tasks. But I'm not sure "success rate" fully captures what we care about with safety. A 10% failure rate in a warehouse is very different from a 10% failure rate in a hospital.

What I do like is that their approach is interpretable. You can actually audit the prompts to see what changed. That's valuable for deployment in regulated environments, and it's something most black-box approaches can't offer.

So where does this leave us?

Both papers are chipping away at the same wall from different angles. LMP says: give language models a better interface to motion control. LAD-VF says: add a verification layer to catch mistakes before they happen.

Neither is a complete solution. We're still far from robots that can reliably execute arbitrary natural language instructions in unstructured environments. But I think these approaches point toward something useful, that the path forward might not be bigger models, but better abstractions.

The gap between "understanding what to do" and "knowing how to do it" is basically the gap between intelligence and embodiment. And honestly, I'm not sure we've been taking that gap seriously enough.

One thing remains unclear to me: how these approaches compose. Could you use LMP's motion primitives with LAD-VF's verification layer? Neither paper addresses this, and it seems like the obvious next question. Maybe someone's already working on it. I should probably ask around.

Three new papers dropped this week that suggest we've been watching the wrong competition.

Sarah Williams · 12 hours ago · 4 min