Robots Are Getting Better at Finding You When You're Lost. Does Anyone Care?

Their approach uses something called Model Predictive Path Integral control, MPPI, which is a sampling-based method that generates trajectory options and picks the best one by optimizing across long time horizons rather than just the next few seconds. The planner handles the search-and-track balancing act, and then when the boat is actually close enough to grab something, it switches to a simpler pursuit controller for the physical capture. They tested it in actual field trials with a real autonomous surface vehicle, not just simulation, which matters more than people give it credit for.

The results show their planner outperforming the baseline approaches they compared against. What we don't know yet is how it performs in genuinely rough conditions, heavy chop, poor visibility, complex debris fields with hundreds of objects. The field trials appear to have been controlled enough to demonstrate the concept cleanly, but the gap between a clean demo and a Category 4 hurricane aftermath is, well, significant.

Second paper: a framework called GLIDE, short for Guided Long-horizon Integrated Drone Escort, which pairs two UAVs with an unmanned ground vehicle to navigate unknown environments during search and rescue. The full details are on arXiv, version four of the paper, which tells you the team has been iterating.

The role separation here is the clever bit. One drone flies ahead doing victim detection and georeferencing, basically finding people and tagging their location. The second drone scouts the terrain ahead of the ground vehicle's planned route, feeding it traversability data so the UGV doesn't drive into a ditch or a collapsed wall. The ground vehicle, which they've demonstrated using a GEM e6 golf cart (a golf cart, I love this), fuses all that aerial information with its own local sensors and replans continuously as new data comes in.

The hardware demo is real, which again, matters. Simulation ablations are useful for understanding the planning stack in isolation, but the fact that they strapped this together on actual robots and ran it is worth something.

The numbers

Both papers are light on the kind of operational numbers that would tell you whether any of this is ready for prime time. The ASV paper reports outperforming baselines but doesn't give you a crisp figure like "recovered 40% more targets per hour." The GLIDE paper reports improved reach time and navigation safety in their test scenarios, but the test scenarios are controlled enough that extrapolating to a real disaster zone requires some imagination.

This is based on limited data, and I want to be honest about that. What I can say is that the technical approaches in both papers are more sophisticated than the "just fly a drone around and look" approach that dominated early SAR robotics research. The MPPI-based planner is doing something genuinely interesting with long-horizon optimization under uncertainty. The role specialization in GLIDE addresses a real coordination problem that multi-robot systems have struggled with for years.

What remains unclear is the failure mode analysis. How does the MPPI planner behave when its ocean current model is wrong? What happens to the GLIDE framework when one of the UAVs loses comms? These aren't gotcha questions, they're the questions that determine whether a system gets deployed or sits in a lab.

So what

Here's my honest read: this is good, incremental, real science. Not the breathless "robots will replace rescue workers" stuff you see in press releases. These are researchers actually grappling with hard subproblems, building systems that handle specific failure modes, and testing on hardware. That's how progress actually happens, and it's sort of unglamorous, which is probably why it doesn't get the coverage it deserves.

The autonomous surface vehicle work feels closer to near-term deployment to me. Robot boats for ocean cleanup and debris recovery don't require the same public trust infrastructure as, say, a self-driving car. Nobody's inside the robot. If it makes a mistake, it misses some debris. The regulatory pathway is clearer, the failure consequences are lower, and there's genuine commercial interest from environmental cleanup organizations. A system that can autonomously sweep for floating litter or locate a man-overboard situation faster than a human crew can respond? That has real value today, not in ten years.

The aerial-ground SAR framework is more ambitious and, I'd argue, further from deployment. Not because the research is weaker, it's not, but because coordinating multiple heterogeneous robots in an actual disaster zone introduces operational complexity that lab results don't fully capture. You've got communications infrastructure that may be damaged. You've got other responders moving through the same space. You've got victims who may be partially obscured or moving. The young researchers behind GLIDE are clearly talented, and the architecture they've built is smart, but this is a system that needs a lot more field time before I'd want to bet lives on it.

Both efforts represent something I find genuinely encouraging though: a move away from single-robot solutions toward coordinated systems that divide labor intelligently. The GLIDE paper's explicit role separation between the goal-searching UAV and the terrain-scouting UAV is the kind of systems thinking that was mostly absent from early multi-robot research. The ASV paper's hybrid approach, long-horizon MPPI for planning, pure pursuit for capture, shows similar maturity about matching the right tool to each phase of the problem.

What happens next

Call me old-fashioned, but I think the path from here to actual deployed systems runs through a few specific bottlenecks that neither paper fully addresses, and that's not a criticism, it's just where the work is.

First: real-world robustness testing at scale. Both systems need to be broken, deliberately, in field conditions. Not to prove they fail, but to understand how they fail and whether those failures are recoverable.

Second: integration with existing emergency response infrastructure. Rescue coordinators don't want to learn a new system in the middle of a crisis. The robots have to fit into how humans already work, not the other way around.

Third, and this is the one that keeps getting skipped: regulatory clarity. The FAA and equivalent bodies in other countries are still sorting out how autonomous multi-vehicle systems get certified for SAR operations. The technology is outrunning the policy, which I've watched happen in autonomous cars and it doesn't end well for anyone. If you want to argue about this, my email's on the about page.

Both of these research threads are worth watching. The ASV work in particular seems like it could find real-world application in the next few years if someone with operational resources picks it up and stress-tests it properly. The GLIDE framework is probably a five-to-ten year story, depending on how aggressively it gets pushed into field trials.

In the meantime: good work, genuinely. Just don't let anyone tell you it's ready to replace search and rescue teams. It isn't. Not yet.

Justin Ernest built a captive LP network to back Anthropic, Anduril, and SpaceX without ever raising a traditional venture fund. Sound familiar?

Mark Kowalski · Yesterday · 7 min