The problem
Mass and friction are hidden until contact, so the planner acts on a guess.
Proof before action
PhysProbe: robots that touch to learn hidden physics before they act.
AgentGlass builds proof before action for autonomous systems: AI that shows it knows enough before it acts, and keeps the evidence.
94.0%task successin simulation
8.0%catastrophic failuresin simulation
59.3%task success without probingin simulation
+14.1points gained by using what touch revealed (K = 81)in simulation
The problem
Mass and friction can't be seen until contact.
A robot can look at a box from every angle and still not know its mass or friction. Those stay hidden until contact. Planners meet uncertainty with more search, and no amount of thinking reveals how heavy a box is.
The solution
PhysProbe touches before it acts and learns the physics it can't see.
Why it matters
Wrong guesses turn into failed tasks and catastrophic failures.
How it's used
The robot probes before it commits, so the planner plans from what the contact revealed.
How it works
Probe → update belief → act
The robot makes a small, deliberate contact, updates what it believes about the object's hidden physics, then acts.
- ProbeA small, deliberate contact with the object, made to learn, not to finish the task.
- Update beliefThe contact updates what the robot believes about the object's hidden physics, such as its mass and friction.
- ActThe robot plans and acts on the updated belief, not on a guess.
Safety gate
A safety gate checks each action against the belief. An action the belief can't support is blocked.
The problem
More search can't reveal what has never been touched.
The solution
Probe → update belief → act, with a safety gate that blocks actions the belief can't support.
Why it matters
The robot acts on evidence, and is held back when the evidence isn't enough.
How it's used
The planner stays the same. It plans from an updated belief, and the gate decides whether the action goes ahead.
Evidence
In simulation, probing lifts task success and cuts catastrophic failures.
The same planner, run with and without probing, in simulation.
Task success and catastrophic failures, same planner, with and without probing
in simulation- With probing
- Same planner, no probing
Task successhigher is better
Catastrophic failureslower is better
Show as a table
| Metric | With probing | Same planner, no probing |
|---|---|---|
| Task success | 94.0% | 59.3% |
| Catastrophic failures | 8.0% | 27.3% |
Pre-registered studyin simulation
“When Does Contact Beat More Search?”
It asked a plain question: should the robot use what it learns by touching, or discard it and search more? In simulation, using it won at each planning budget tested. More planning compute alone did not close the gap.
+8.0K = 9
+7.3K = 27
+14.1K = 81
Points gained by using what touch revealed, at planning budget K.
Paper in preparation.The problem
Claims that touch helps are easy to make and hard to check.
The solution
A pre-registered study and a head-to-head against the same planner without probing, all in simulation.
Why it matters
The analysis was fixed before the results were in, so the result can't be picked after the fact.
How it's used
Read every number here as a simulation result. The paper is in preparation.
Who it's for
For robotics engineers and world-model researchers.
Robotics engineers putting manipulation into warehouses, labs and factories
PhysProbe gives the robot a way to learn an object's mass and friction before it commits, and a safety gate that blocks actions the belief can't support.
World-model researchers who need to know when their model is guessing
PhysProbe keeps an explicit belief about hidden physics and acts on it through a gate. When the belief is thin, the gate blocks the action, so a guess is visible instead of silent.
The problem
Both act on models of objects that have never been touched.
The solution
A probe that measures what the model guessed, and a gate that blocks actions the belief can't support.
Why it matters
Knowing when a model is guessing is what separates a safe stop from a catastrophic failure.
How it's used
Robotics engineers add the probe step before the planner commits. World-model researchers read the belief to see when the model is guessing.
Software agents
See AgentGlass for software agents →The same idea for software agents
AgentGlass applies proof before action to AI software agents too: release gates and signed evidence for AI agents.
Sidra Miconi, PhD · @SidraMiconi
I build proof before action because an autonomous system should show what it knows before it moves.
Working on manipulation or world models?
Questions about the probe, the gate or the study: I want to hear them.
Get in touch