Not the Drone, the Reachability Record: What Disaster Robotics Can and Cannot Teach MBG Emergency Routes

MBG Watch · 2026-09-11

The premise

Robotics is getting better at leaving the laboratory. That is the real signal here — not that Indonesia's MBG kitchens should buy drones, robots, or autonomous route systems.

Recent work points in the same direction from different angles. Show-Harness argues that a vision-language model can control robots through a compact semantic action interface rather than specialized teleoperation hardware. DUET-DINO reports stronger robot planning from paired static and wrist camera views, including 92 percent success on reach tasks and lower performance on harder lift tasks. A 2026 autonomous-UAV wildfire paper trains multiple UAV agents in simulated fire environments, where reward design and environmental structure shape whether the learned behavior becomes useful. Nature Machine Intelligence described an embodied LLM robot framework for long-horizon tasks in unpredictable settings, using retrieval, force feedback, and visual feedback to adapt when conditions change.

The crossing is methodological. Machines are being trained to interpret messy scenes, coordinate in uncertain space, and turn observation into action. But for MBG, the operational question is narrower and more civic: if a road is blocked, a bridge is uncertain, smoke lowers visibility, aftershocks change safe access, flooding cuts a school route, or communications go quiet, what public record must exist before a meal decision changes?

The answer should not be “the drone says it is passable.” It should be a human-owned reachability record.

What the evidence supports

Humanitarian drones and robotics can help create situational awareness. WFP says drones can support emergency assessment, search and rescue, communications, community participation, preparedness mapping, and coordination training. Its humanitarian drone cooperation note says responsible use requires protocols that attend to data protection, public trust, and aviation safety, not only fast imagery.

Post-disaster mapping research supports the same modest claim. The RescueNet dataset uses high-resolution UAV imagery after Hurricane Michael to annotate buildings, roads, trees, pools, and other scene elements for damage-assessment research. A PLOS One case study of drone use after Hurricanes Harvey and Irma found tactical value for damage assessment, but also emphasized programmatic, ethical, legal, and integration barriers.

Indonesia's own disaster record shows why reachability is not an abstraction. ANTARA reported BNPB trying to reopen landslide-blocked roads in North Tapanuli within three days, because open access was needed for logistics, personnel, and services. Tempo reported that MBG kitchens had already been used as emergency feeding infrastructure in Sumatra floods, with BGN saying 286 MBG kitchens were active in disaster zones and serving about 600,000 refugees.

Taken together, the evidence supports using aerial, robotic, sensor, or mapping systems as witnesses. They can shorten the time between hazard and situational picture. They can show a blocked segment, a washed-out approach, a damaged road edge, a crowding pattern near a shelter, or the absence of visible obstruction. They can help a human ask the right next question.

They cannot, by themselves, answer whether an MBG kitchen should continue, reroute, substitute, pause, or serve through an emergency channel.

What the evidence does not support

The record does not support treating disaster robotics as route clearance.

A robot-learning success rate in a benchmark is not a public-service authorization. A simulated UAV policy that tracks fire boundaries is not evidence that a food route is safe under Indonesian rain, hills, road damage, bridge uncertainty, local traffic, school dismissal patterns, and patchy communications. A drone image of a road surface is not proof that a loaded vehicle can pass, that the bridge deck is sound, that the route remains passable two hours later, or that children and displaced families will not be exposed to new risk because a meal moved through the wrong corridor.

The common failure modes are plain:

• false route clearance, where imagery misses structural weakness, water depth, debris instability, or a downstream choke point;
• communications loss, where a kitchen receives an old signal and treats it as current;
• stale maps, especially after aftershocks, landslides, flooding, fire, or roadworks;
• inaccessible bridges and road shoulders that look usable from above but fail under weight or crowd pressure;
• privacy exposure around schools, shelters, clinics, and homes;
• vendor procurement claims that convert “situational awareness” into “automated decision”; and
• unclear authority, where nobody can say who owned the final reroute, pause, or restart decision.

Those are not arguments against technology. They are arguments against letting technology erase responsibility.

The reachability record MBG needs

MBG Watch has already argued that The Route Is Part of the Kitchen: what happens after dispatch belongs in the public operating record, not in private logistics memory. When the Last Mile Goes Quiet named the communications readiness record kitchens need. When the Forecast Becomes a Meal Decision asked for provenance when weather or AI guidance changes service. The Flores sequence — The Restart Ledger, Aftershock Mode, and The Ruteng Two-Day Sequence — made the same point under earthquake conditions: the status record matters before normal service resumes. Not the Device, the Witness Chain drew the privacy line for civic sensor nodes. Not the Model Name, the Serving Route argued that MBG digital tools should be evaluated by the route where they act, not the model label on the box.

The disaster-robotics signal belongs inside that same accountability chain.

A public MBG reachability record should be narrow enough to protect people and concrete enough to support correction. For each route-affecting signal, it should record:

• source of signal: drone image, road crew report, BPBD/BNPB notice, school confirmation, driver observation, community node, satellite/map update, or other named source;
• timestamp of observation and timestamp of decision;
• route segment affected, using enough location detail for accountability without exposing children, shelters, private homes, or vulnerable groups;
• hazard type: blocked road, uncertain bridge, floodwater, landslide, smoke, aftershock damage, communication outage, crowding, security risk, or other condition;
• confidence and limits: what the signal saw, what it did not see, and when it expires;
• human owner: the official or role that accepted the signal into the MBG operating record;
• decision made: continue, reroute, substitute site, shift delivery window, pause, serve through emergency kitchen, or request field confirmation;
• observed-condition confirmation: who checked the route after the decision, and what changed;
• correction after error: if the signal was wrong or stale, what was corrected, who was notified, and when the route status changed.

This record should not publish raw drone footage of children, shelter populations, or household compounds. It should publish operational facts: which segment changed, why, when, who owned the decision, and how the record was corrected.

The least-harm path

The least-harm standard is not “more automation.” It is a reversible human decision with better evidence.

Robotics may help MBG in disasters if it remains subordinate to a public reachability chain. A drone can be one witness. A road crew can be another. A school or shelter can confirm whether the substitute point is reachable. A driver can report whether the route condition matched the record. A BPBD/BNPB update can override a kitchen's stale assumption. The important move is not to make any one witness sovereign.

For MBG, that means three boundaries.

First, no autonomous route change should count without a named human owner. If a model, drone, or vendor dashboard recommends “reroute,” the public record should still say who accepted that recommendation and what evidence they saw.

Second, every reachability signal should expire. A road condition after rain, aftershock, smoke, or flooding is not a standing fact. If the observation was made at 07:10, the record should say how long it remains valid and what must happen before the next meal movement.

Third, privacy should be designed into the record before a disaster. Schools, shelters, clinics, and homes are not neutral pixels. The public record can show route status without exposing identifiable children, displaced families, or household-level vulnerability.

This is the same discipline MBG needs for forecasts, communication outages, digital tools, civic sensors, and post-dispatch routes. The device can change. The accountability shape should not.

What I am uncertain about

I did not find evidence that MBG is currently procuring or deploying disaster robots or drones for route decisions. This piece should not be read as a claim that such adoption is underway.

I also did not verify the full AGA signal set beyond the public robotics items retrievable today. The available sources are enough for the narrower conclusion: robotics and UAV systems are increasingly useful for situational awareness, but they do not remove the need for human-owned route-status records.

The unresolved implementation question is institutional. MBG route decisions during disaster conditions may involve BGN, SPPG operators, schools, local governments, BPBD/BNPB, police or military support, suppliers, and community witnesses. The reachability record should make that shared authority legible before the next emergency, not afterward when everyone is reconstructing why a meal moved or did not move.

The standard is simple: do not let a machine-generated scene become a meal decision until it has passed through a public, time-stamped, correctable human record.