Two New AI Planning Papers Want Robots to Think Before They Act. Here's Why That Matters on the Floor.

The second paper introduces PearlVLA, which tackles a different but related headache. Modern vision-language-action models, the kind that can take a camera image and a text instruction and output a robot action, tend to face a nasty tradeoff. Fast systems are dumb; thoughtful systems are slow. Too slow for real-time control.

PearlVLA's answer is to move the deliberation into what they call latent space, basically the internal compressed representation inside the neural network, rather than making the robot write out reasoning steps in plain text or generate explicit visual subgoals. The system refines its action plan through multiple internal passes before committing, and it does this quickly enough to still be useful. They test it on LIBERO, a robotic manipulation benchmark, and report state-of-the-art results among current methods.

What I find interesting here is the underlying philosophy. The robot isn't just reacting. It's sort of running a quick internal simulation of what might happen before it moves. That's closer to how a skilled human operator thinks, actually. I've watched good technicians pause for half a second before a tricky pick, running through the motion in their head. This is attempting to formalize that.

Now, the caveats. Both papers are preprints on arXiv, not peer-reviewed yet. The results are promising but lab-scale. Neither paper talks much about what happens when you put this stuff in a real warehouse with inconsistent lighting, dust on the lens, and a forklift driver who keeps moving the staging area. That gap between benchmark and deployment is where ideas go to die, and I've seen it happen more times than I can count.

It also remains unclear how either system handles novel objects it hasn't seen during training. That's the question that always comes up in industrial settings and it's the one these abstracts don't really answer.

Still, I think the direction is right. The field has been stuck in a mode where you either have a fast dumb system or a slow smart one. Anything that moves the intelligence into the planning layer without killing latency is worth watching. Whether COMET or PearlVLA specifically make it into a commercial product is another question entirely. But someone building the next generation of warehouse picking systems is probably reading these papers right now.

This raises questions about... well, multiple things. Mostly around how long the gap between academic robotics and deployable industrial robotics is going to stay as wide as it currently is. That's a longer conversation.

Robert "Bob" Macintosh · Yesterday · 4 min