The POMDP Problem Nobody Wants to Talk About Is Finally Getting Solved

The catch, and there's always a catch, is that they're solving for a "reference-based optimal solution" rather than the true POMDP optimum. The paper admits this. Whether that gap matters in practice remains unclear. I'd want to see this tested on something messier than their benchmark problems.

VOPP: Parallelization done right

The second solver, VOPP (Vectorized Online POMDP Planner), takes a different angle. The problem with parallelizing POMDP solvers has always been synchronization. You've got processes that need to wait on each other, which kills your speedup. I've seen teams throw GPUs at planning problems and get basically nothing back because the architecture didn't fit.

VOPP sidesteps this by, well, restructuring the whole computation. They represent everything as tensors and vectorize all the planning steps. No dependencies between processes. arXiv claims 20x efficiency gains over existing parallel solvers and, more impressively, they match state-of-the-art sequential solvers with 1000x less planning budget.

I'll be honest: I don't fully understand the math behind their "analytically solving part of the optimization component." But the results suggest something real is happening here.

Why I'm cautiously optimistic

I called my old colleague Marcus at Siemens last week. He's still doing logistics automation, and I asked him what planning approaches they're actually using in production. His answer was depressing: mostly reactive systems with some model-predictive control for the easy cases. POMDPs are still too slow, too brittle, too hard to tune.

These two solvers might change that. Maybe. The 3000-step horizon from ROP-RAS3 would cover most warehouse scenarios I can think of. The parallelization in VOPP means you could actually run this on edge hardware without waiting three seconds for a decision.

But I've been burned before. Academic benchmarks and real warehouses are different beasts. The papers don't mention sensor noise, communication dropouts, or what happens when a human walks into the planned path. They show a "physical robot demonstration" but the details are sparse.

What I'd want to see next

Someone needs to test these on a real multi-robot warehouse scenario. Not a demo cell, a real facility with 50+ AMRs, human workers, and all the chaos that implies. I'm too old and too retired to do it myself, but if anyone from the RDL Lab or the VOPP team is reading this: call me. I know people who'd give you access to test environments.

The other thing I'd want is failure mode analysis. When these solvers get it wrong, how wrong do they get it? A planner that's 95% optimal but occasionally drives into a wall is worse than one that's 85% optimal and fails gracefully.

For now, I'm filing this under "promising but unproven." Which, in my experience, is better than most academic robotics work. At least these folks are trying to solve a problem I actually had.

Robert "Bob" Macintosh · 13 hours ago · 3 min