A steel apple falls. The robot still trusts the picture.
ReactHuman executes robot plans in simulation instead of grading text alone. Seven multimodal models mishandled about one in three sudden hazards in the authors’ tests.
DAILY RESEARCH NOTE · PREPRINT
An apple can look safe and still be made of steel. An anvil can look threatening and be foam. ReactHuman puts that mismatch in front of a multimodal model, asks for an immediate robot action, and executes the plan in a 240 Hz rigid-body simulator. I like the uncomfortable part: the answer has to arrive somewhere, not merely sound right.
QUICK ANSWER
Short answer
ReactHuman executes robot plans in simulation instead of grading text alone. Seven multimodal models mishandled about one in three sudden hazards in the authors' tests.
A steel apple, a foam anvil, and one second to act
- The preprint describes more than 1,000 reproducible scenes across 17 event families. Ground truth comes from simulation rather than manual labels. Each committed plan is executed, so the evaluator can see whether a hand reaches the interception point, whether the robot avoids danger, and whether the response matches the object's actual physics. Five metrics cover reasonableness, safety, and physical grounding.
- Primary source: ReactHuman preprint
The result that matters
The authors tested seven representative multimodal large language models. They report that the models mishandled roughly one in three hazards. Failures included fixed action patterns, trusting visual appearance over motion, and missing an interception point by meter-scale distances even when the chosen action type was correct. The paper reports no consistent improvement with model scale. These are preprint results from the authors' harness, not independently replicated field performance.
Why text-only grading misses the failure
Catch the plate is a verb, not a trajectory. A useful reaction needs timing, reach, mass, velocity, collision geometry, and some account of the cost of being wrong. Executing the plan makes those choices visible. The adversarial objects ask whether the model follows physical cues or falls back on anvil means heavy.
A minimal interception check
The original code below shows one boundary the benchmark makes visible: the contact point must be reachable before impact. It is explanatory code, not the ReactHuman implementation, a benchmark reproduction, or a robot controller.
def reachable_before_impact(hand, target, t_impact, max_speed):
distance = ((target.x-hand.x)**2 + (target.y-hand.y)**2 +
(target.z-hand.z)**2) ** 0.5
arrival = distance / max_speed
return {"reachable": arrival < t_impact,
"arrival_s": arrival, "impact_s": t_impact}This ReactHuman sketch explains a boundary; it is not measured evidence, the source implementation, or production-ready software.
Concrete replication test
For a concrete proposed test, keep geometry and motion fixed. Compare an ordinary apple, a steel apple, a foam anvil, and a real anvil across repeated camera sequences. Log the action, contact point, arrival time, collision, and injury proxy. Changing mass should matter; changing texture alone should not steer the plan. These are test criteria, not new results.
Jack's analysis
My favorite feature is the commitment: an answer becomes an event. Next I would test uncertainty and abstention. A household robot may need to retreat, warn, or stop rather than attempt a brittle catch. Count perception, inference, control, and actuator delay in latency. A plan that succeeds only while the scene waits for it is not yet a safe reflex.
Limits and uncertainty
The paper is an arXiv preprint submitted on 9 September 2026. Simulation gives repeatability, but contact, sensing noise, motors, human movement, and damage are messier in homes. The published abstract and first-party summary do not establish that the full code, model prompts, and all scenarios are independently reproducible today. "One in three" summarizes the authors' evaluated set; it is not a universal robot failure rate.
Takeaway
Do not ask only whether a robot chose the right verb. Execute the trajectory. Measure where the hand arrives, when it arrives, and what it hits. Then change appearance without changing physics, and physics without changing appearance. A model that cannot survive those swaps should not be trusted near a falling knife.
FAQ
What is the short answer?
ReactHuman executes robot plans in simulation instead of grading text alone. Seven multimodal models mishandled about one in three sudden hazards in the authors' tests.
What evidence does this article use?
For ReactHuman, use the source links to check the reported claims. This article keeps 3 named sources in that trail. My proposed tests are not additional evidence that the approach works.
What should readers keep separate?
In this reading of ReactHuman, reported claims stay with their sources. The suggested workflow and my judgments are analysis; unresolved questions remain unresolved.